# Igor Barani: every published piece > Igor Barani works on typography, design and artificial intelligence. He writes mathematics in which every step carries its reason. His day job is radiation oncology. 13 pieces, newest first. Each is the Markdown copy served beside its page. Index: https://barani.org/llms.txt ======================================================================== URL: https://barani.org/writing/what-eighth-grade-algebra-replaces/ Markdown: https://barani.org/writing/what-eighth-grade-algebra-replaces/index.md Published: 2026-11-19 Modified: 2026-11-19 Form: Essay Topic: Mathematics teaching Series: How children learn mathematics, part 3 of 4 ======================================================================== # What does eighth-grade algebra replace? Put Algebra I on an eighth grader’s schedule. What comes off? Igor Barani · 19 November 2026 · 40 min read Mathematics teaching · Essay · How children learn mathematics, part 3 of 4 · https://barani.org/writing/what-eighth-grade-algebra-replaces/ Your eighth grader has a place in Algebra I and a place in band. Both fit, the counselor explains, if your child drops Math 8. Would you sign? I would keep eighth-grade algebra open. Before I signed, I would ask the algebra teacher to show me one piece of work. Somewhere between seventh grade and ninth, a student should be able to explain why a straight line has the same slope everywhere. Where will my child learn to do that? If the teacher can open the book to the lesson and show me what students write, I am much closer to saying yes. If the lesson lives in the class my child is about to drop, we have found a problem while it is still easy to fix. Then I would hand the schedule to the eighth grader. Band has a chair with someone’s name on it, too. A child who has practiced an instrument for three years has information about the price of this offer that the math department does not have. Algebra I in eighth grade takes the place of something. Taken in the regular math period, it replaces Math 8, the course in which California’s standards place that lesson. Taken as a second math class, it replaces an elective or a support period. An offer of eighth-grade algebra should say what it replaces, and four written answers would do it. What changes on this student’s timetable? Where will the displaced mathematics be taught? Who checks the fit, and when? What can the student take next fall? One district has now written the second answer down. San Francisco’s guidance to families for 2026–27 says that a student who opts out of Math 8 skips eight units of mathematics, linear equations among them. Those units, “if skipped, would not be taught at another time” [1]. California law does not ask a district for that sentence. The placement law the state passed in 2015 covers students entering ninth grade and leaves the eighth-grade offer to each district [2]. The state’s course files show how far the course fell after 2013, when the state stopped expecting it of eighth graders and left the choice to districts [3, 4]. On census day in 2012–13, 64 percent of California eighth graders were enrolled in Algebra I or a later course. Six years later 20 percent were. In San Francisco’s district-run schools the share went from 91 percent to 2 percent in one year, and then to none [5]. Early algebra has good grounds behind it. In a randomized experiment in Maine and Vermont, schools offered a real algebra course to students they judged ready. It improved their algebra and roughly doubled their chance of following an advanced sequence in high school. Caution has good grounds too. In another district, pushing a much wider group into eighth-grade algebra made some of those students less likely to pass geometry on time. Both results can be true. To use either one, a family needs to know what the students would otherwise have studied and what happened the following September [6, 7]. ## The lesson that can fall between two classrooms Draw a straight line rising across squared paper. From one point on it, go two squares right and three up to meet it again. Go four right from the same starting point and it takes six up. A student who divides 3 by 2 and 6 by 4 gets the right answer twice. Now choose two points much closer together, halfway up the page, where neither sits on a grid intersection. Why must the answer come out the same? The horizontal and vertical steps make a right triangle. Make a second triangle anywhere else along the line. Its horizontal side is parallel to the first horizontal side, so it meets the sloping line at the same angle. Each triangle also has a right angle. Two matching angles settle the shape. One triangle is a scaled copy of the other, and whatever factor stretches the horizontal side stretches the vertical side with it. Dividing the rise by the run cancels that factor. You can now put the ruler down. The other pairs have already been accounted for. > **Figure 1.** The same rising line in three panels of squared paper, with a right triangle under it in each. The first two triangles start at the same grid corner and run 2 squares, then 4. The last pair of points starts elsewhere on the line, halfway up, with neither end on a grid corner. Moving the start and shrinking the triangle leave the ratio unchanged, at 3 ÷ 2, 6 ÷ 4 and 1.2 ÷ 0.8. All three panels use equal horizontal and vertical unit scales. Constructed from the example. > > What the figure shows: First of three panels showing the same rising straight line on squared paper. A right triangle under the line starts at a grid corner, runs 2 squares right and rises 3. Rise divided by run is 3 divided by 2, which is 1.5. Second panel, the same line. From the same corner the triangle runs 4 squares right and rises 6. Rise divided by run is 6 divided by 4, which is 1.5. Third panel, the same line. A small triangle starts elsewhere, halfway up, between grid corners, running 0.8 of a square and rising 1.2. Rise divided by run is 1.2 divided by 0.8, which is 1.5, as in the other two. That little proof does more work than a page of slope calculations. Suppose the line crosses the vertical axis at 5. Going two units right adds 3 to the height. Going one unit right adds 1½. Going any distance x to the right adds 1½ times x, so the height is y = 5 + 1½x. Going left works the same way, with x negative. The 5 tells us where the line starts at x = 0. The 1½ tells us how it changes. The triangle connects a picture to an equation the student will use for years. California’s grade-eight standard 8.EE.6 explicitly asks students to make that connection. Algebra I has a cousin of it. Its standard F-LE.1a asks students to prove that linear functions grow by equal differences over equal intervals. That proof usually starts from the equation, and the grade-eight lesson is where the equation comes from. [8] Try changing the picture. If the line falls as you move right, its vertical change has a minus sign. A horizontal line has zero rise for any nonzero run. On a vertical line the run between any two points is zero, so there is no real number you can obtain by dividing rise by run. Those cases help a teacher hear what a student means by “slope.” A student can make a good argument in ordinary language or with a drawing. Saying “angle-angle similarity” is optional. Explaining what stays in proportion is the work. A school could teach the whole argument beautifully in Algebra I. It could also teach it beautifully in Math 8 and begin Algebra I with “Remember how we find the slope?” Each teacher would be making a sensible decision for the students the course was designed to receive. Only the student traveling a different route would encounter the gap. A course catalog with “linear equations” under both headings would give the family no warning. Researchers have looked inside this apparently ordinary word. Courtney Nagle and colleagues followed slope through seven successive textbooks in one series. It appeared in all seven, with different emphasis on slope as a ratio, a constant in an equation, or a description of steepness. In September 2026 Marios Pittalis reported on six lessons taught by two eighth-grade teachers in Cyprus. The teachers spent most of their time on slope as a ratio read from a graph, and their 116 students showed a limited understanding of the idea. Neither followed children who skipped Math 8. They do explain why asking for a piece of student work is worth the trouble [9, 10]. Slope happens to fit on a napkin. A district should give the rest of the course the same courtesy. Where do students use the Pythagorean theorem to find a distance between points? Where do they examine a scatterplot and decide whether a straight line is a sensible description of the data? These are different mathematical acts [8]. ## What San Francisco put in writing San Francisco Unified has printed its answer. Beginning in 2026–27, all of its middle and K–8 schools offer Algebra I to eighth graders, under a policy the Board of Education approved on March 24, 2026. At 19 schools a student can take it as a second class beside Math 8, which takes the week’s math from 220 minutes to 440 in the district’s proposed schedule. Two more schools are piloting a sequence that covers Math 6, Math 7, Math 8 and Algebra I in three years [11, 12, 13]. The guidance the district gives eighth-grade families states the rule for the second class. Enrollment in Algebra I is automatic for a student with a proficient score on the sixth-grade state math test and a C or better in fall seventh-grade math. The family can opt out, and other students can opt in [1]. The same page tells a family what the other route costs. “To ‘opt out’ of Math 8 is to skip eight different units of mathematics,” it says. The units, it goes on, “have topics that are foundational to future math courses and, if skipped, would not be taught at another time” [1]. The topics it names include linear functions and linear equations, geometric transformations and the Pythagorean theorem. That is the second of the four answers, written by the district that makes the offer. Linear equations are where California’s grade-eight standards put the lesson on the napkin [8]. The district’s pages, as they stood on November 19, 2026, disagree on who may take that route. The family guidance lets a student skip Math 8 with the top level on the sixth-grade state test or on the winter STAR test in seventh grade. The pathways page, updated on July 16, 2026, asks for the top level on both the fall and the winter STAR tests. It also asks for a rating of “secure” in every seventh-grade standard. Both require a counselor meeting and a parent’s signed consent [1, 13]. The guidance also asks one comparison to carry too much. It tells families that Stanford researchers “studied the effect of skipping Math 8” and found more students repeating Algebra I in ninth grade. The figures are 19.2 percent, against 9.4 percent for students who took both courses [1]. They come from two different sets of pilot schools, and the students eligible to appear in them had reached algebra in quite different ways. A family cannot use the ten-point gap as an estimate of how much a second period would lower one child’s risk of repeating [14]. In the fall of 2013, 91 percent of eighth graders in its district-run schools were enrolled in Algebra I or a later course. In 2014 the district moved Algebra I to ninth grade [5, 15]. Two years later it called the new ninth-grade course “a substantial departure from the course previously taught as Algebra 1” [16]. “Many topics previously covered in eighth-grade algebra are now covered in ninth grade,” the same release said. So, by the district’s 2016 account, the Algebra I taught since the change shares a name with the Algebra I of 2013 and differs from it in content. In March 2024 the city’s ballot carried Proposition G, a declaration of policy on offering Algebra I to eighth graders. The City Attorney’s memorandum on the results directs the Board of Supervisors, which does not run the schools, to take any action within its power to carry it into effect [17]. The district ran a pilot in 2024–25 [18]. On May 29, 2025, lawyers for students and two community groups who had petitioned under the state’s placement law reported an agreement. By their account the district agreed to drop its Math Validation Test, set up an appeal process and reinstate Algebra I in eighth grade [19]. That account is the petitioners’ counsel’s, and neither the court record nor the agreement was retrieved. The district’s own announcement came on June 16, ahead of a board vote the next day. It said students no longer had to pass the test to take geometry in ninth grade, a change it called a clarification of its current practice. “Algebra 1 options will be available to all 8th grade students,” it added, and it mentions no petition [20]. ## What California’s own files show The state publishes a file of every class section in its public schools. Each row gives the course, the grade, and the race or ethnicity of the students enrolled on census day in early October. The published files run from 2012–13 to 2018–19. The count made for this essay takes eighth-grade seats in Algebra I, in the integrated course that takes its place, or in a later course such as geometry. It divides them by eighth-grade enrollment on the same day [5]. In 2012–13 the share was 64 percent, and in 2018–19 it was 20. Most of the fall came in two years. Leaving out Los Angeles Unified, whose course codes are inconsistent in the middle years, the share was 52 percent in 2013–14 and 25 percent in 2014–15. Elizabeth Huffaker, one of the two researchers who evaluated San Francisco’s pilot, published the statewide fall in May 2026 from the same records. She counts roughly one hundred districts that had offered the course widely in 2012 and had 2 percent or fewer of their eighth graders in it by 2018 [21]. In those years the state changed three things, and then each district chose. In January 2013 the State Board of Education took Algebra I out of the eighth-grade standards. In March it ended the accountability rule that had marked down a school’s results when an eighth grader took general math instead [3, 4]. The state’s end-of-course Algebra I test ended after spring 2013. Districts that sharply limited the course went from 64 percent to 10, and the districts Huffaker compares them with still went from 67 to 46. So the course fell even where districts kept it on offer, and fell furthest where they chose to limit it [21]. The files also give enrollment by group. No published series of eighth-grade enrollment by race for these years was found, and Huffaker reports race only for later outcomes. Outside Los Angeles Unified, the share of Asian eighth graders in Algebra I or a later course went from 79 percent to 43. Among white students it went from 66 to 24, among Hispanic or Latino students from 60 to 14, and among African American students from 57 to 12 (Figure 2). The gap between Asian and African American students widened from about 23 points to about 30 [5, 21]. > **Figure 2.** Share of California eighth graders enrolled in Algebra I or a later course on census day, by race and ethnicity, 2012–13 to 2018–19. All students fell from 64 percent to 20. Asian students fell from 79 to 43, white students from 66 to 24, Hispanic or Latino students from 60 to 14, and African American students from 57 to 12. The dotted line is San Francisco Unified’s district-run schools, at 91 percent in 2013–14, 2 the next year and zero by 2017–18. Every line counts seats on one October day, capped at each school’s enrollment, and the state lines leave out Los Angeles Unified. In 2012–13 the state’s accountability rules expected Algebra I of eighth graders, and by 2018–19 it was an accelerated course. The figure does not show completion, learning or cause. California Department of Education course and census enrollment files [5]. > > What the figure shows: Line chart of the percent of California eighth graders enrolled in Algebra I or a later course on census day, 2012–13 to 2018–19, in all public schools except Los Angeles Unified. All students fell from 64 percent to 20. Asian students fell from 79 to 43, white students from 66 to 24, Hispanic or Latino students from 60 to 14, and African American students from 57 to 12. More than half of each fall came in one year, between 2013–14 and 2014–15. The gap between Asian and African American students widened from about 23 points to about 30. In 2012–13 enrollment in Algebra I alone was close to equal across the four groups, between 53 and 57 percent. A dotted line shows San Francisco Unified’s district-run schools: 93 percent in 2012–13, 91 in 2013–14, 2 in 2014–15 and zero from 2017–18. Every line counts course seats on census day, capped at each school’s enrollment. In 2012–13 the state’s accountability rules expected Algebra I of eighth graders, and by 2018–19 it was an accelerated course. The figure does not show completion, learning or cause. Those lines record what came off timetables and whose changed most. They do not measure harm. In 2012–13 the state’s rules expected Algebra I of every eighth grader, and enrollment in Algebra I alone was close to equal across groups. It stood at 57 percent of Asian eighth graders, 56 percent of white, 57 percent of Hispanic or Latino and 53 percent of African American. The gap sat in the courses above Algebra I. Twenty-two percent of Asian eighth graders were in geometry or beyond, against under 4 percent of African American eighth graders. By 2018–19 Algebra I had become the accelerated course, so one line joins two different placements [5, 22]. The files also count seats, not students, and October enrollment, not completion. A year after 64 percent of eighth graders sat in Algebra I or beyond, geometry and later courses held 38 percent of ninth-grade math seats. That suggests many took Algebra I again [5]. The outcomes measured so far point the other way. Huffaker estimates that where districts pulled algebra back, the gap in later advanced courses between underrepresented students and others narrowed. It narrowed by “leveling down rather than lifting up,” she writes, because students already well represented lost more. She finds no detectable effect on eleventh-grade math scores in her main estimate, though her estimates lean negative, and fewer ninth graders repeating Algebra I. Enrollment in precalculus and beyond fell by roughly two to three percentage points in districts that sharply limited the course. That fall was concentrated in small and mid-size districts, and weighted by the number of students it is about one point and cannot be told from zero [21]. The files stop at 2018–19. District leaders interviewed for a 2025 PACE study reported eighth-grade algebra shares close to the 2018 figures, a mean of about 20 percent [21]. Los Angeles Unified’s course codes in the middle years do not match where its ninth graders went next, so the yearly figures leave it out. With it, the first and last figures are unchanged [5, 23]. On census day in 2013–14, 3,258 of the 3,593 eighth graders in San Francisco’s district-run schools were in Algebra I or a later course. The next fall 65 of 3,497 were, and by 2017–18 none. Before the change every racial group in those schools was above 74 percent, so within one year the course ended for nearly everyone [5]. A study of successive cohorts later traced San Francisco’s students through high school. Participation in Advanced Placement mathematics initially fell by six percentage points, largely through AP Calculus. The decline diminished as students made greater use of ways to accelerate later. Large racial and ethnic differences in advanced course-taking persisted. Other things changed across those cohorts, so this is a history of what happened through the reform, with limits on assigning cause. It suggests a question for any school that promises another chance. How many students who started on the regular route used it last year [15]? On federal civil rights data, which count Algebra I alone, California ranked about 46th of 51 in 2015–16, at 17 percent against 25 nationally. States label courses differently, so a rank is a rough guide [24]. A child can lose a place in eighth-grade algebra long before the course request form comes home. The school may never offer it. In the federal 2021–22 Civil Rights Data Collection, 39 percent of the roughly 31,900 public middle schools serving grade seven or eight did not offer Algebra I. Only 48 percent of public high schools offered calculus. Each school counts once in those figures, regardless of enrollment. They are also two ends of the same practical question. An early start buys a place in a later class only if someone offers that class, or the school provides a workable route to it [25]. In the same collection’s school file, 52 percent of California’s 3,212 schools serving those grades reported no Algebra I class for them. The share was 31 percent in Texas and 21 percent in North Carolina. Most of the California schools without a class also enroll the earliest grades. Among the state’s separate middle schools, about 36 percent reported none [24]. ## What the law asks, and of whom California’s legislature took up placement in 2015. The Mathematics Placement Act found that successful students, disproportionately students of color, were being “held back to repeat 8th grade mathematics coursework” [2]. An Assembly committee’s analysis of the bill cited a study of nine districts. In it, nearly 65 percent of students placed in algebra in eighth grade were placed in algebra again in ninth [26]. The law they passed, Education Code section 51224.7, binds one step. A district serving students entering ninth grade must adopt a “fair, objective, and transparent” placement policy. The policy must include a checkpoint within the first month of school and offer “clear and timely recourse” to a family that questions a placement. It must also require the district to examine its placement data each year, by race, ethnicity, gender and socioeconomic background, and to report the results to its governing board. For the step from elementary to middle school, the same statute says a district “may” adopt such a policy. The committee’s staff had found no statewide data on placement at that earlier step and recommended the permissive wording. The offer at the counselor’s desk in seventh or eighth grade falls under no requirement in the section [2, 26]. The statute leaves the checking to each district’s own governing board. In one county, residents went to court. In February 2023 the Santa Clara County Superior Court, in case 21CV385694, found Palo Alto Unified School District in violation of the section’s first four requirements. Ninth-grade placement there was “based solely on each student’s eighth grade math class,” the order states. On the yearly examination of data it records that “No evidence of actual data collection or reporting to the governing board was presented in the briefs or at the hearing” [27]. The court ordered the district to submit a compliant policy within 30 days and set a compliance hearing for March 13, 2023. The order was read from a copy that has not been checked against the court’s record, and the district’s present policy was not checked. The petition came from private individuals, and no state agency was a party. The agreement the petitioners’ lawyers report in San Francisco in 2025 also began with a petition from students and community groups [19]. Even where algebra appears in the catalog, asking for it can be a separate hurdle. Wake County, North Carolina, used students’ prior test records to estimate the probability of passing the Algebra I end-of-course exam. A predicted chance of 70 percent triggered a recommendation for accelerated mathematics. Placement became more closely related to prior academic performance and less closely related to race and income. This does not make a cutoff a perfect measure of readiness. It makes the invitation less dependent on a parent knowing whom to call [28]. In 2018 North Carolina wrote an eighth-grade rule into statute. It applies where a district offers advanced courses. A seventh grader who scores at the highest level on the state’s end-of-grade math test “shall be enrolled in a high school level mathematics course in eighth grade,” the statute says. The school can remove that student only with a parent’s written consent. The same section orders a yearly report of how many students were eligible and how many were placed [29]. The report due in December 2025 counts 14,164 eligible eighth graders and 13,673 placed. Among Black eighth graders, 830 of 899 were placed, or 92 percent. Across grades six through twelve the state placed 65,206 of 70,763 eligible students, and two districts placed fewer than 15 percent of theirs [30]. California’s law asks for no such public count. Texas moved the invitation back to sixth grade. Its statute directs districts and open-enrollment charter schools to develop a program leading toward Algebra I in eighth grade. Sixth graders who qualify through a fifth-grade math score or a qualifying local measure enter advanced mathematics automatically, with a parent opt-out. One route into automatic enrollment is a score in the top 40 percent statewide on the fifth-grade mathematics assessment. The state also publishes tables showing where content from the ordinary sequence goes in the compacted one. A parent can ask to see the row for a particular idea. A seventh grader arriving from another district can use the same map to identify the work already taught [31, 32]. ## The best case for saying yes Imagine a small school with one eighth-grade math teacher and three students ready for algebra. A friend in a larger town can take geometry next fall. The three students would have to wait a year because too few classmates need a separate section. In this case, insisting on the regular course also has an educational cost. It asks students to spend a scarce year on material they may already command. A federal experiment in Maine and Vermont came close to that situation. Sixty-eight schools, most of them rural, did not ordinarily offer a stand-alone eighth-grade Algebra I class. Before the researchers assigned schools by lottery, school staff identified students they judged ready. In the schools offered the new course, ready students could take algebra online during the school day with a remote teacher and an adult in the room. In the comparison schools, usual practice continued. Some of those schools still found a way to provide formal algebra to about one in five of their ready students. The experiment tested the effect of making an opportunity reliably available [6, 33]. The offer raised algebra achievement by 0.39 standard deviations on the study’s algebra assessment. A standard deviation expresses a difference relative to the spread of scores. The researchers also tested mathematics outside algebra. The estimated difference there favored the offered group, by about 0.14 standard deviations, but was too imprecise to distinguish from zero. By the start of tenth grade the predicted proportion following an advanced sequence was 51 percent in the schools offered online algebra and 26 percent in the comparison schools. To count, a student had to complete a full-year course above Algebra I in ninth grade with at least a C and be enrolled in the next course for tenth grade. Finishing one year and taking a seat in the next is a reasonably demanding meaning of “advanced” [6]. Those results put a limit on how much caution a school should ask a prepared family to exercise. The algebra gain was measured directly, the original offer was randomized, and the opportunity persisted past the first transcript entry. One check in the report cuts the other way. When the researchers used the study’s own pretest as the starting measure, the algebra gain was 0.29, and the course-taking difference no longer met the study’s threshold for significance. The course also came with an instructor, scheduled time and local supervision. Sending home a login and asking a thirteen-year-old to find spare time would change the opportunity considerably [6, 33]. The experiment took place in 2008–09. Its usual eighth-grade courses already contained considerable algebra, and the present California standard on similar triangles had yet to be adopted. The nonalgebra test result is reassuring for the students and material studied. It cannot promise that a modern replacement course preserves every topic. Suppose a school can point to the lessons that matter, and the receiving high school will honor the credit. Then keeping band and taking algebra is a choice with serious evidence behind it [6, 8]. Charlotte-Mecklenburg supplies the counterexample. Beginning in 2002–03, the district rapidly raised eighth-grade algebra participation among moderately performing students from about half to 85 percent. It subsequently reversed course. Following cohorts exposed to the change, Charles Clotfelter, Helen Ladd and Jacob Vigdor found worse Algebra I end-of-course performance among affected students and less success reaching geometry on time. Teachers with weaker measured qualifications were assigned to algebra in the first year, but that accounted for only a small portion of the adverse effect. The school system had changed who entered, how quickly it expanded, and whom it had available to teach. The study cannot divide the result neatly among those changes [7]. Instruction can change the outcome of a placement decision. In a separate randomized study, Huffaker and Thomas Dee followed ninth graders whose prior scores would ordinarily have placed them in remedial pre-algebra. Their district combined access to Algebra I with teacher training and time to plan, and the targeted students did better in mathematics two years later, with no measured loss for classmates. What will the teacher be able to do differently on Monday when this wider group arrives [34]? ## What is the extra year worth? Write five course names on five slips of paper: Algebra I, geometry, Algebra II, precalculus and calculus. In a school that teaches one each year in that order, beginning in ninth grade leaves four high-school years for five slips. Beginning in eighth makes all five fit. That is the strongest simple reason parents care. A student can decide at sixteen whether to take calculus if the choice still exists at sixteen. No one needs to know at twelve whether the student will want to be an engineer. The conclusion depends on the slips. An integrated sequence has different course names. A school might offer concurrent courses, a compacted sequence later on, or a college course that fits senior year. A program that requires a paid summer course, a long bus ride or an adult free at noon creates a different opportunity from one built into the school day. And the University of California, for example, specifies three years of college-preparatory mathematics covering algebra and geometry and recommends a fourth. Its general subject requirement does not stipulate high-school calculus. Families can value the option without mistaking one particular course for a universal admissions requirement [13, 35]. California researchers Andrew McEachin, Thurston Domina and Andrew Penner studied students near placement cutoffs at 510 middle schools. For students whose placement changed at the threshold, early algebra increased advanced math enrollment by about 30 percentage points in ninth grade and 16 in eleventh. The tenth-grade gain on the state’s high-school exit exam was much smaller, about 0.05 standard deviations. The benefits were larger at schools that set a higher bar, and the authors note a possible trade-off between wider access and later success. The students were enrolled before today’s Math 8 existed. The finding nevertheless makes sense of the slips on the table. A relatively small difference on one academic test can travel alongside a substantial difference in what a school permits a student to take [36]. Two studies followed students further. In Los Angeles, Miller and Phillips matched students within schools. Those who started advanced math in seventh grade, most of them Latino and low-income, were about ten percentage points more likely to take calculus or higher mathematics in high school. Students were not assigned by lottery, so unmeasured differences could explain part of the advantage [23]. In a large Florida district, David Card and Laura Giuliano compared students near the score cutoff for a program one year ahead of the district’s other advanced track. Entering it increased completion of calculus by an estimated 9.5 percentage points. Effects on overall college entry were close to zero, and the effects, the abstract says, “vary widely by non-cognitive skill and gender” [37]. It would stretch the result to promise a family that one year of early algebra increases a child’s probability of graduating from college. At a school-board meeting, any one of these outcomes might be called success. A policy can change what a student knows, which room the student is allowed to enter, and what that student eventually chooses to do there. Each can matter. Each needs its own evidence. A 2026 Regional Educational Laboratory scan of ten studies, which its authors call limited, found a more consistent association with advanced course-taking than with achievement [38]. ## San Francisco measured two ways to make room San Francisco tried five arrangements in 2024–25, among them an online class and a summer course. Two were measured side by side. Six schools offered Algebra I as a second math class alongside Math 8. Three others put nearly the whole eighth grade into Algebra I in the regular math period. At those three schools, 639 of 675 students enrolled in algebra. Some had access to a supporting math lab. The district selected pilot sites in light of staffing and student readiness. Each school began with its own timetable, teachers and students [14, 18]. Even the invitation changed between design and practice. The six schools were supposed to enroll qualified students automatically. According to Elizabeth Huffaker and Thomas Dee, none carried that out as planned. In the researchers’ panel, 360 students took the elective and 653 schoolmates did not. The student who signs up for a second period may have a family who knows the course sequence, or room on the schedule. A teacher may say, quite personally, “You should do this.” Those are plausible reasons for a difference between groups. The enrollment record cannot tell us which one mattered for a particular child [14]. Huffaker and Dee compared changes in scores over time, checked earlier trends, and examined English alongside mathematics. They presented the early results in March 2026, in a slide deck that has not been through peer review. Figure 3 shows the estimates from that presentation. Positive values mean greater gains than the relevant comparison group. The units in the figure tell us how far apart the changes were. They cannot be read as percentages of questions answered correctly. The same researchers and their colleagues produced the statewide estimates, the San Francisco history and the ninth-grade trial, so much of the California evidence rests on one group [14]. > **Figure 3.** Estimated differences in grade-eight state-test growth in San Francisco’s 2024–25 pilot, in standard deviations, for math and for English. Each group uses a different comparison. Algebra I alongside Math 8, at six schools, compares participants with schoolmates who did not enroll: math +0.235, English −0.009. Algebra I in the regular period, at three schools, compares pilot schools with schools outside the pilot: math −0.062, English −0.064. Estimates from slide 22 of the March 2026 presentation [14], with 95 percent intervals calculated from its standard errors. The interval for the extra-class math estimate runs from +0.174 to +0.296. > > What the figure shows: Two separate comparisons from San Francisco’s 2024–25 pilot, in standard deviations, with 95 percent intervals. Algebra I alongside Math 8, six schools, participants against schoolmates who did not enroll: math plus 0.235, interval plus 0.174 to plus 0.296; English minus 0.009, interval minus 0.076 to plus 0.059. Algebra I in the regular period, three schools, pilot schools against schools outside the pilot: math minus 0.062, interval minus 0.115 to minus 0.009; English minus 0.064, interval minus 0.123 to minus 0.005. The two groups use different comparisons. The first group is encouraging. Students who took both courses gained substantially more on the math test than the schoolmates used for comparison. English was essentially flat. The second group is harder to interpret. The modest decline in mathematics was almost exactly matched by the decline in English, and both declines are distinguishable from zero, narrowly. A schoolwide change affecting both subjects is one possible explanation. The researchers estimated the difference between the math and English changes. It was +0.244 standard deviations for the extra-class model and +0.002 for the three-school model. On a district assessment the next fall, students who had taken both courses still showed a smaller positive math difference. English on the same chart reads slightly below zero for both arrangements. English is a useful check, but it cannot tell us whether anything changed inside the mathematics classroom. Subtracting the two math entries would give the appearance of a head-to-head trial that San Francisco never ran [14]. What did the second period buy? It gave a student more mathematics and preserved Math 8 in the year the state tested grade-eight content. Either feature could help. The state assessment samples a year’s curriculum across different kinds of tasks. Its overall score can improve while one particular explanation remains shaky. An Algebra I test can reveal gains that a grade-eight test barely touches. The question “Did the students learn more?” deserves a follow-up. More of which mathematics, measured when [8, 14, 39]? Then there is the page a band director might turn to first. In the broader fall scheduling data, the 391 students taking the extra math class averaged 0.99 freely chosen electives. The 940 who did not take it averaged 1.65. The difference is about two-thirds of an elective per student. When the researchers also counted academic support periods in their larger elective category, the averages were 1.07 and 2.12. The detail that surprised me is on the right of the same slide. Read from the bars, roughly three in ten algebra takers were enrolled in music, compared with roughly one in four of the other students. Visual arts went the other way. The groups had different tastes, eligibility and schedules before any comparison began. The chart cannot tell us who would have taken band under a different timetable. It does tell us why the letter should name the particular class each student would give up [14]. In Houston, a lottery expanded access to arts experiences at 42 elementary and middle schools, and Daniel Bowen and Brian Kisida found improvements in writing and fewer disciplinary infractions. The lottery did not ask eighth graders to choose between band and algebra. Applying its average to the child with the clarinet would be as careless as assigning that child San Francisco’s average math gain. A family can value playing in an ensemble for the music itself [40]. One more scheduling question belongs here. If a student qualifies for algebra but the second period conflicts with English language development or another essential service, what offer has that student received? The family choosing between band and algebra has two choices it can reasonably entertain. A family that needs a particular support period may have only one. The San Francisco deck notes that only about 2 percent of algebra takers were English learners. Its records do not establish that algebra displaced a required language service. A list of students who declined, by itself, records all these situations in the same column [14]. ## What Olivia’s transcript couldn’t tell her In Tacoma, Washington, a student given the pseudonym Olivia told researchers that she had taken Algebra II in ninth grade and found herself in it again in tenth. She thought she had earned As. She had not asked to repeat it. She was also willing to take it again because the earlier course had been online and she felt unsure how much she had learned. Her account is about Algebra II, one student, and the interruption of schooling during the pandemic. It captures something no placement spreadsheet could have supplied. The student had separate answers to “Did you pass?”, “Did you choose this?” and “Did you understand it?” [41] Tacoma has run Algebra for All in eighth grade since at least 2013. In an August 2026 federal report, researchers reconstructed the paths of 2,778 students who were seniors in 2022–23 or 2023–24 and had been in the district since at least seventh grade. Eighty-seven percent earned at least one semester of Algebra I credit in eighth grade. In grades nine and ten, 34 percent of the full sample followed Algebra I then geometry. Another 33 percent followed geometry then Algebra II. The first percentage includes students with different amounts of eighth-grade credit [41]. The divergence mattered years later. In the first group, 23 percent enrolled in advanced math by twelfth grade. In the second, 67 percent did. Those are unadjusted rates. Students reaching geometry in ninth grade differed in preparation and circumstances from students taking Algebra I that year. Moving the same child from one room to the other would not necessarily deliver a 44-point improvement. When the researchers accounted for prior achievement, the difference narrowed only slightly. The interviews offer a more useful question. Teachers described different local rules for repetition. Students and counselors described uneven influence over placement. Ask who takes responsibility when the eighth-grade teacher’s recommendation and the ninth-grade schedule disagree [41]. For students close to Wake County’s cutoff, a later study found that acceleration increased the likelihood of passing ninth-grade geometry by about 33 percentage points. Middle-school test gains were less clear, and grades could fall as students took harder courses. The advantage in advanced course-taking narrowed as students moved through high school. A school celebrating a long list of acceptances in August should look for those same names again when the geometry class meets, and again when precalculus begins. A student can earn a lower grade in a harder class and still get somewhere the old timetable would never have allowed [23, 42]. San Francisco provides an example of a written answer. Its published guidance says a student who completes district coursework covering the Algebra I standards before ninth grade and earns at least a C in Algebra I will enter geometry. For an outside course, it asks for a syllabus or equivalent. For an online course, it also specifies accreditation and eligibility under California’s a–g college-preparatory requirements. Ninth graders receive a placement check within the first month. That is the kind of information a family should have before giving up an elective or paying for a summer course [13]. ## The strongest case against asking The best objection to all of this comes from people who want more children in algebra, as I do. No study has shown the gap. At San Francisco’s three single-period schools, the math change net of the English change was +0.002 standard deviations. In Maine and Vermont the students who skipped the usual course scored a little higher in other mathematics, though the difference could be chance. Meanwhile every added step loses families. Wake County’s rule took one step away, the parent’s request. A counselor meeting, a consent form and a letter with four answers would add three [6, 14, 28]. The objection is right that every added step loses some families. A written answer adds no step. North Carolina and Texas both enroll by default. North Carolina still counts every placement, and Texas publishes where each topic goes. The letter is the school’s to write, and it records a timetable decision the school has already made. As for the missing study, the single-period result is a difference between schools on an early, unreviewed estimate, and the Maine and Vermont courses predate the present standard. The district that ran the pilot now tells families the skipped units are not taught again [1, 30, 32]. A second objection comes from the other direction, and it has more evidence behind it. California once pushed most eighth graders into algebra, and learning did not follow enrollment. In 2008 the state moved to make Algebra I its accountability benchmark for eighth-grade mathematics. Thurston Domina and colleagues found that enrolling more eighth graders in the course had negative average effects on tenth-grade achievement, driven by large districts [22]. Charlotte found much the same in one district [7]. When the State Board took Algebra I out of the eighth-grade standards in January 2013, EdSource reported that 60 percent of minority eighth graders in the course had not tested proficient. The board’s vice president said the intent was “to deepen rigor” [3]. San Francisco reported in 2017 that its Algebra 1 repeat rate had fallen from 40 percent to 8 after its change. Tom Loveless, citing a parent group’s analysis of district records, answered that the district had also changed the rules for passing the course [43, 44]. Huffaker’s estimates, given above, support it [21]. The state’s 2023 mathematics framework, which leaves the design of courses and pathways to local districts, accepts that some students will be ready for Algebra I in eighth grade. It says successful acceleration needs a strong foundation. It recalls that many students in the algebra-for-all years repeated the course in ninth grade, and it describes a high-school route to calculus as an alternative [45]. Waiting also gives every eighth grader the full Math 8 course, slope lesson included. Most of this case I accept, because more eighth graders in algebra did not mean more algebra learned, and Figure 2 does not measure harm. The decision at the counselor’s desk remains. Test scores did not rise when districts pulled the course back. By Huffaker’s estimate calculus enrollment fell by about 15 percent of its starting level, and underrepresented students saw no offsetting gains [21]. Broad acceleration cost some students one thing and broad delay cost others another, so each family is left asking what its own child gives up. The four answers work for a school opening the course and for one closing it. ## Four answers worth putting in writing **What changes on this student’s timetable?** Give the family the two schedules side by side. Name the class that moves, who teaches the additional mathematics, and what happens to band, language development or another support period. Let the student say which loss would matter. A student who would give up band happily and a student who would miss it every morning can both have understood the offer perfectly. **Where will the displaced mathematics be taught?** A one-page map can give the course, unit and a representative assignment for the content that moves. Show the slope explanation. Show one geometry problem and one investigation of data. In a compressed sequence, tell a transfer student what the class has already done and how the school will help the student join. A child should not have to repeat two years or skip an idea because the family moved across a district line. **Who checks the fit, and when?** Name a teacher or counselor and a date while a change of schedule is still workable. A first-quarter review might look at an unfamiliar problem and ask the student to explain one decision. If the difficulty is a missing prerequisite, the teacher can teach that prerequisite and check again. If the difficulty is persistent across topics or the timetable is costing too much, the school and family can consider a change. An early review is fairer than discovering in June that the student spent a year in the wrong room. **What can the student take next fall?** Name the receiving course, the credit or grade required and what happens if the high school has questions. Give the regular Math 8 family the next chance to accelerate, whether it runs in the school day, what it costs, and who has used it. REL Northwest, working with Tacoma, has produced a family guide and a school decision template for the transition to ninth grade. A district that wants to improve its letter already has somewhere concrete to begin [46]. A school could answer the opening family in a few lines. Band stays on the schedule. Algebra I meets during the Math 8 period. In September the algebra class uses similar triangles to explain constant slope. In October the two math teachers will review the first unit with your child. Here is the receiving high school’s rule for entering geometry next fall. This is an example. The dates, assignments and credit rule have to come from the school that sends the letter. Three public bodies can answer for many schools at once. I have sent these questions to no one, and I print them so that a parent, a board member or a reporter can ask them. A California school board can say how many seventh graders met its rule last year, how many were offered a place and how many enrolled, by group. North Carolina publishes that count for every district. San Francisco’s board can say which of its two tests for skipping Math 8 applies, how many students skipped it this fall, and which class each two-course student gave up. A board adding sections can say who teaches them, because in Charlotte the first year’s new sections went to teachers with weaker measured qualifications. Any board can say whether the yearly examination of placement data that section 51224.7 requires has reached it, and whether the public may read it [1, 2, 7, 13, 30]. The State Board of Education and the California Department of Education can say whether anyone collects districts’ placement policies, or the reports their boards are owed. The section names no state agency for that work [2]. The Department can publish course enrollment by grade again. Its downloadable course files, as retrieved for this essay, stop at 2018–19, so no one can count from them what San Francisco, or any other district, has restored since [5]. The Legislature’s education committees wrote section 51224.7 eleven years ago to bind the step into ninth grade. They can extend it to eighth grade, with North Carolina’s count of students eligible and students placed, by district and by group [2, 29]. ## Check Start with the worry about slope. Give both groups a rising line with two pairs of points and ask for a justification. A week later, offer a falling line in a different setting. Accept an equally sound explanation expressed through coordinates, proportional changes or a diagram. A student who succeeds on the unfamiliar example has shown something worth keeping. A student who struggles has shown a teacher something worth teaching. If students who skipped Math 8 explain the falling line as well as students who took it, my worry about that school is wrong. Two children at a kitchen table cannot tell us which schedule caused either outcome. They may have come to the table with different knowledge, teachers or opportunities to practice. A district can do more. Before placement, record preparation and a short sample of students’ reasoning. Keep counts of who qualified, who received an offer, who accepted, who encountered a timetable conflict, who finished and who advanced. Give both routes a common algebra measure and a separate sample of the grade-eight work the district means to preserve. Report counts alongside percentages. A disappearing student can make either course look better than it was. Then follow the cohort into ninth grade. Separate students who earned full credit from those who earned partial credit or none. Among repeaters, distinguish a student who asked for another year, a student who needed a missing topic, and a student whose receiving school would not accept a course. If there are fewer second-period seats than equally eligible volunteers, a fair lottery can answer a harder question. Follow everyone offered a seat and everyone assigned the available alternative, including students who later switch. The California count can be checked the same way, because the files are public and the references give the course codes and the method [5]. If the state’s records since 2018–19 show districts restoring eighth-grade algebra evenly across groups, Figure 2 describes a past that has been repaired. If San Francisco reports that students who meet its rule are placed at the same rate in every group, the question of who gets the invitation is answered for that city. On the skipped lessons San Francisco has already answered in writing. What remains to learn there is how many students take that route, and how they fare. No study found for this essay tested the slope argument in students who skipped Math 8. I would recommend a single-period route more widely if students entering it kept their grade-eight understanding, gained in algebra and advanced the next fall. I would favor an additional period for more students if it improved both kinds of mathematics enough to justify what those students had to give up. My confidence is highest about the modest proposal that survives either result. The people who build the timetable are in the best position to explain it. I write and will sell the With Every Reason books, elementary mathematics books with a kit of materials, published by my company, KAGU, LLC. They are still being written and are not yet on sale. I will sell them to families and to schools, including through district adoptions. An essay that asks schools for a map of where each idea is taught runs with my interest. I would hold my own pages to the test in this Check. Back at the counselor’s desk, the family may decide that algebra and band belong together. Another family may choose both math classes. Ask the eighth grader what matters, let the teachers open their books, and have the high school say what comes next. Then hand over the pen. The decision has a counterpart in [a lesson on adding a half and a third](https://barani.org/writing/when-to-show-the-method/). What I would show next depends on which ideas the child can already use. ## Keep An offer of eighth-grade algebra should say what it replaces. ## Sources 1. San Francisco Unified School District. 8th grade math course enrollment family guidance, 2026–27 school year. Linked from the district’s secondary mathematics pathways page. Retrieved November 19, 2026. [Family guidance](https://docs.google.com/document/d/1IoxY4hdffsbzAqm9utBzCRbgafjvz9Q1xLW0JO2_iiw/edit). 2. California Legislature. Senate Bill 359, California Mathematics Placement Act of 2015, enrolled text, adding Education Code section 51224.7. Section 1(c) holds the finding quoted. Subdivision (b) sets the ninth-grade requirements and subdivision (c) the permissive clause. [Enrolled bill](https://www.leginfo.ca.gov/pub/15-16/bill/sen/sb_0351-0400/sb_359_bill_20150908_enrolled.htm). 3. EdSource. It’s final. State Board shifts policy on eighth grade Algebra. January 2013. Reports the board’s 10–0 vote, the proficiency figures and the vice president’s remark. [News report](https://edsource.org/2013/its-final-state-board-shifts-policy-on-eighth-grade-algebra/25672). 4. EdSource. State Board eliminates incentives to offer Algebra in 8th grade. March 2013. Describes how an eighth grader’s General Math result had been marked down one level in a school’s accountability score. [News report](https://edsource.org/2013/state-board-eliminates-incentives-to-offer-algebra-in-8th-grade/28525). 5. California Department of Education. Course enrollment files, 2012–13 to 2018–19, with the Department’s course code list and its census-day enrollment files by school, grade, and race or ethnicity. Method: grade 8 seats on census day in Algebra I (codes 2403, 2437, 2467), Integrated Mathematics I (2425, 2440) or a later course (geometry, Algebra II, Integrated Mathematics II to IV, precalculus, statistics, calculus and International Baccalaureate courses: codes 2404, 2407–2415, 2417, 2422, 2426, 2427, 2430, 2438, 2439, 2441–2445, 2460–2463, 2468, 2473, 2480, 2481, 2483–2485 and 2490), divided by census-day grade 8 enrollment. Every figure, for all students and for each group, caps a school’s seats at its enrollment of that group. The files count seats in course sections, so they record enrollment and neither students nor completion. Two-year algebra courses and the code “accelerated math” (2423), whose content the code list does not define, are left out. Counting 2423 would put the 2018–19 figure for African American students at 18 percent. Figures by year and by group exclude Los Angeles Unified. The whole-state figures for 2012–13 and 2018–19 are 64 and 20 percent with or without it. Asian excludes Filipino students, as in the Department’s categories. San Francisco figures cover district-run schools. Los Angeles Unified reported 60 percent of its eighth graders in Algebra I or beyond in 2013–14, 88 percent the next year and 19 percent in 2017–18, yet after the 88 percent year most of its ninth graders were in Algebra I and few in geometry. The district was not asked about its coding. Three smaller districts with swings of 40 points that later reversed are left in, and dropping any one of the fifteen largest remaining districts moves the first and last figures by less than half a point. The count was run twice, by separate programs. Figures are rounded to whole numbers because reasonable choices of course codes move the 2012–13 figures by two to four points. [Course enrollment files](https://www.cde.ca.gov/ds/ad/filesassign.asp). 6. Jessica B. Heppen and colleagues. Access to Algebra I. The effects of online mathematics for grade 8 students. NCEE 2012-4021. Regional Educational Laboratory Northeast and Islands, December 2011. Randomized school offer. The executive summary and Tables 4-1, 4-2 and 4-3 distinguish algebra, general mathematics and later course-taking. [Full federal evaluation](https://files.eric.ed.gov/fulltext/ED527394.pdf). 7. Charles T. Clotfelter, Helen F. Ladd and Jacob L. Vigdor. The aftermath of accelerating algebra. Evidence from a district policy initiative. National Bureau of Economic Research Working Paper 18161, June 2012. Published in the Journal of Human Resources 50(1):159–188, 2015. [Charlotte-Mecklenburg study](https://www.nber.org/papers/w18161). 8. California Department of Education. California Common Core State Standards. Mathematics. Adopted August 2010, modified January 2013. Grade 8, standards 8.EE.6, 8.G.8 and 8.SP.1–2, and Algebra I, standard F-LE.1a. The examples in this essay illustrate the standards. [State standards](https://www.cde.ca.gov/be/st/ss/documents/ccssmathstandardaug2013.pdf). 9. Courtney Nagle, Deborah Moore-Russo and Thomas Fisher. A study of slope explanations across a standards-based textbook series. International Journal for Mathematics Teaching and Learning 23(1):25–45, 2022. Analysis of seven books from one series. [Article](https://doi.org/10.4256/ijmtl.v23i1.403). 10. Marios Pittalis. Slope conceptualizations addressed by Grade 8 teachers and their impact on student understanding. Journal of Mathematics Teacher Education, published online September 11, 2026. Six lessons, two teachers and 116 students in Cyprus. The abstract reports the lesson observations and the student assessment. [Article](https://doi.org/10.1007/s10857-026-09783-w). 11. San Francisco Unified School District. Board of Education approves revised math placement policy. Board action of March 24, 2026, posted March 26 and updated March 29. [Approved policy](https://www.sfusd.edu/announcements/2026-03-29-board-education-approves-revised-math-placement-policy). 12. San Francisco Unified School District. Shaping math learning together. Algebra I in eighth grade update. March 19, 2026. Published before the board’s vote. [Weekly minutes and school options](https://www.sfusd.edu/announcements/2026-03-19-shaping-math-learning-together-algebra-i-eighth-grade-update). 13. San Francisco Unified School District. SFUSD secondary mathematics pathways. Family guidance for 2026–27, updated July 16, 2026, covering Math 8 exemptions, transfer credit, ninth-grade placement and the checkpoint. The page links the family guidance for eighth grade. [District guidance](https://www.sfusd.edu/departments/mathematics-department-page/sfusd-secondary-mathematics-pathways). 14. Elizabeth Huffaker and Thomas S. Dee. Algebra I pilot research methodology and findings. Presentation of early results at a San Francisco Unified School District webinar, March 23, 2026. Slides 4, 8, 10, 14, 16, 20, 22, 24, 30–32 and 37. Not peer reviewed. [Announcement and recording](https://www.sfusd.edu/announcements/2026-03-24-algebra-i-pilot-research-methodology-and-findings-webinar-recording-available). [Slide deck](https://drive.google.com/file/d/1UXLzO44Pq7OOj7ErzOUnFND-aDGmrykv/view). 15. Elizabeth Huffaker, Sarah Novicoff and Thomas S. Dee. Ahead of the game? Course-taking patterns under a math pathways reform. Educational Researcher 54(2):91–102, 2025. Successive San Francisco cohorts across the change to ninth-grade algebra. [Published study](https://doi.org/10.3102/0013189X241309642). 16. San Francisco Unified School District. San Francisco Board of Education adopts math placement policy. Press release, June 15, 2016. [Press release](https://www.sfusd.edu/about-sfusd/sfusd-news/press-releases/2016-06-15-san-francisco-board-education-adopts-math-placement-policy). 17. City and County of San Francisco, Office of the City Attorney. Memorandum on the March 5, 2024 election results. Lists Proposition G, “Offering Algebra 1 to Eighth Graders,” under declarations of policy, and directs the Board of Supervisors to take any action within its power. [Memorandum](https://media.api.sf.gov/documents/Memorandum-re.-March-5-2024-Election-Results.pdf). 18. San Francisco Unified School District. SFUSD announces pilot schools for Algebra 1 in 8th grade in 2024–25. March 15, 2024. [Press release](https://www.sfusd.edu/about-sfusd/sfusd-news/press-releases/2024-03-15-sfusd-announces-pilot-schools-algebra-1-8th-grade-2024-25). 19. Rains Lucia Stern St. Phalle & Silver. SFUSD to halt controversial math placement test, comply with state law. Statement of May 29, 2025, by the law firm that represented the petitioners. An interested source. The petition, the court record and the agreement were not retrieved. [Counsel’s statement](https://www.rlslawyers.com/sfusd-to-halt-contoversial-math-placement-test-comply-with-state-law/). 20. San Francisco Unified School District. SFUSD math placement policy update. Announcement of June 16, 2025, before the board’s vote of June 17. [Announcement](https://www.sfusd.edu/announcements/2025-06-16-sfusd-math-placement-policy-update). 21. Elizabeth Huffaker. Adoption windows and reform. California’s math pathways in the Common Core era. Getting Down to Facts and Policy Analysis for California Education, May 2026. The source for the statewide decline and for the estimates of what followed. Footnote 3 defines the count of districts. [Report](https://gettingdowntofacts.com/sites/default/files/Adoption%20Windows%20and%20Reform_%20California%E2%80%99s%20Math%20Pathways%20in%20the%20Post-Common%20Core%20Era.pdf). 22. Thurston Domina, Andrew McEachin, Andrew Penner and Emily Penner. Aiming high and falling short. California’s eighth-grade algebra-for-all effort. Educational Evaluation and Policy Analysis 37(3):275–295, 2015. District-level changes in enrollment and a tenth-grade achievement measure. Read in abstract. [Study](https://doi.org/10.3102/0162373714543685). 23. Carrie E. Miller and Meredith Phillips. Long-term consequences of early access to educational opportunity. American Educational Research Journal 62(3):651–686, 2025. Matched comparison within schools. College enrollment covers one cohort. The article also summarizes the published Wake County follow-up and, on page 3 of the working-paper version, the decline in Los Angeles. [Study](https://doi.org/10.3102/00028312251331023). 24. U.S. Department of Education, Office for Civil Rights. Civil Rights Data Collection, public-use school files for 2013–14, 2015–16, 2017–18, 2020–21 and 2021–22. Calculated for this essay: eighth-grade Algebra I enrollment by state, divided by grade 8 membership from the Common Core of Data, as served by the Urban Institute’s Education Data Portal, and the share of schools serving grade 7 or 8, justice facilities excluded, that reported no Algebra I class for those grades in 2021–22. [Data files](https://civilrightsdata.ed.gov/data). [Enrollment data](https://educationdata.urban.org/documentation/schools.html). 25. U.S. Department of Education, Office for Civil Rights. 2021–22 Civil Rights Data Collection. A first look. January 2025, page 7. The Algebra I and calculus measures use different school-level denominators. [Federal report](https://www.ed.gov/media/document/2021-22-crdc-first-look-report-109194.pdf). 26. California Assembly Committee on Education. Bill analysis of Senate Bill 359, hearing of July 15, 2015. The analysis reports the 65 percent figure from a 2010 study of nine districts and 1,700 students, which was not retrieved for this essay. [Committee analysis](https://www.leginfo.ca.gov/pub/15-16/bill/sen/sb_0351-0400/sb_359_cfa_20150715_104048_asm_comm.html). 27. Superior Court of California, County of Santa Clara. Order after hearing on petition for writ of mandate, case no. 21CV385694, filed February 6, 2023. The order was prepared by the petitioners’ counsel and signed by the court. Read from a copy of the filed order that has not been checked against the court’s record. 28. Shaun M. Dougherty, Joshua S. Goodman, Darryl V. Hill, Erica G. Litke and Lindsay C. Page. Middle school math acceleration and equitable access to eighth-grade algebra. Educational Evaluation and Policy Analysis 37(1 suppl):80S–101S, 2015. [Placement study](https://doi.org/10.3102/0162373715576076). 29. North Carolina General Statutes, section 115C-81.36, Advanced courses in mathematics, subsections (a) to (c). Subsection (a) asks boards to offer advanced courses “when practicable,” and the placement rule in (b) applies where they are offered. Session Law 2018-32. [Statute](https://www.ncleg.gov/EnactedLegislation/Statutes/HTML/BySection/Chapter_115C/GS_115C-81.36.html). 30. North Carolina Department of Public Instruction. Report to the North Carolina General Assembly. Advanced courses in mathematics, G.S. 115C-81.36. Due December 15, 2025. Data for the 2025–26 school year. Summary of findings and Tables 1, 2 and 5. [State report](https://webservices.ncleg.gov/ViewDocSiteFile/104658). 31. Texas Legislature. Senate Bill 2124, 88th Legislature, enrolled text, 2023. Texas Education Code section 28.029. [Enrolled bill](https://capitol.texas.gov/tlodocs/88R/billtext/html/SB02124F.htm). 32. Texas Education Agency. Middle school advanced mathematics program, with its frequently asked questions and crosswalks. The questions page gives the statewide top 40 percent. [Program guidance](https://tea.texas.gov/academics/subject-areas/mathematics/middle-school-advanced-mathematics-program). [Questions page](https://tea.texas.gov/academics/subject-areas/mathematics/middle-school-advanced-mathematics-program-faq). 33. What Works Clearinghouse. Quick review of Access to Algebra I. March 2012. An independent assessment of the Maine and Vermont trial. [Independent review](https://ies.ed.gov/ncee/wwc/Docs/QuickReview/algebra_032712.pdf). 34. Thomas S. Dee and Elizabeth Huffaker. Accelerating opportunity. The effects of instructionally supported detracking. American Educational Research Journal 63(2):307–350, 2026. A randomized ninth-grade study. [Study](https://doi.org/10.3102/00028312251408539). 35. University of California. First-year subject requirement (A–G), area C, mathematics. [Admissions requirement](https://admission.universityofcalifornia.edu/admission-requirements/first-year-requirements/subject-requirement-a-g.html). 36. Andrew McEachin, Thurston Domina and Andrew Penner. Heterogeneous effects of early algebra across California middle schools. Journal of Policy Analysis and Management 39(3):772–800, 2020. Placement thresholds at 510 schools identify effects for students near those cutoffs. The abstract reports larger benefits at schools with higher thresholds. [Study](https://doi.org/10.1002/pam.22202). 37. David Card and Laura Giuliano. Targeted acceleration in middle school math. Impacts on college entry, degree completion, and STEM. National Bureau of Economic Research Working Paper 35847, October 2026. Regression discontinuity at a fifth-grade cutoff. Tables 4 to 6, and the abstract on how the effects vary. A working paper. [Working paper](https://www.nber.org/papers/w35847). 38. Regional Educational Laboratory Central. Summary of key findings on outcomes related to mathematics pathways for junior high. Ask an Expert handout, June 2026. A limited scan of ten studies. [Handout](https://ies.ed.gov/sites/default/files/rel-central/document/2026/06/REL-CE-MathPathways-AAE-handout.pdf). 39. Smarter Balanced Assessment Consortium. 2024–25 Summative Technical Report, Chapter 4, Test Design. [Technical report](https://technicalreports.smarterbalanced.org/2024-25_summative-report/_book/test-design.html). 40. Daniel H. Bowen and Brian Kisida. Investigating the causal effects of arts education. Journal of Policy Analysis and Management 42(3):624–647, 2023. A school-level lottery among 42 elementary and middle schools in Houston. [Study](https://doi.org/10.1002/pam.22449). 41. Noman Khanani, Lilla Pivnick, Edrick Willie Sabalburo and Steven Tedeschi. Grade 9–10 math coursetaking sequences, advanced math outcomes, and students’ course decisions in Tacoma Public Schools. REL 2026-018. Regional Educational Laboratory Northwest, August 2026. Figures 1 and 5 and the interviews. Olivia’s account is on printed page 11, and the adjusted comparison on page 7. The study does not identify the causal effect of repetition. [Federal report](https://files.eric.ed.gov/fulltext/ED682470.pdf). 42. Shaun M. Dougherty, Joshua S. Goodman, Darryl V. Hill, Erica G. Litke and Lindsay C. Page. Objective course placement and college readiness. Evidence from targeted middle school math acceleration. National Bureau of Economic Research Working Paper 21395, July 2015, revised April 2017. Published in Economics of Education Review 58:141–161, 2017. Table 5 reports 0.328 for passing geometry by ninth grade. [Working paper](https://www.nber.org/papers/w21395). 43. San Francisco Unified School District. Policy shifts in math show promise. Press release, September 14, 2017. The district’s own evaluation. [Press release](https://www.sfusd.edu/about-sfusd/sfusd-news/press-releases/2017-09-14-policy-shifts-math-show-promise). 44. Tom Loveless. San Francisco’s detracking experiment. Education Next, March 29, 2022. Reports an analysis of district records by the parent group Families for San Francisco. [Article](https://www.educationnext.org/san-franciscos-detracking-experiment/). 45. California Department of Education. Mathematics framework for California public schools, kindergarten through grade twelve. Adopted by the State Board of Education, November 2023. Chapter 8, the high-school chapter, on acceleration in middle school, and the framework’s summary, on local autonomy. [Chapter 8](https://www.cde.ca.gov/CI/ma/cf/documents/mathframeworkch8.pdf). [Summary](https://www.cde.ca.gov/CI/ma/cf/documents/mathfwsummary.pdf). 46. Lilla Pivnick. Grade 9 math placement. A family guide and school decisionmaking template. Regional Educational Laboratory Northwest, May 2026. [Family guide and template](https://ies.ed.gov/use-work/resource-library/resource/fact-sheetinfographicfaq/grade-9-math-placement-family-guide-and-school-decisionmaking-template). --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/when-to-show-the-method/ Markdown: https://barani.org/writing/when-to-show-the-method/index.md Published: 2026-11-19 Modified: 2026-11-19 Form: Essay Topic: Mathematics teaching Series: How children learn mathematics, part 4 of 4 ======================================================================== # When should a teacher show a child how to add fractions? A child adds a half and a third and gets less than a half. Igor Barani · 19 November 2026 · 27 min read Mathematics teaching · Essay · How children learn mathematics, part 4 of 4 · https://barani.org/writing/when-to-show-the-method/ A child writes 1/2 + 1/3 = 2/5. What would you say next? I would cover the answer with my hand. Put a strip of paper on the table and call its length one. Mark halfway. “If you start here and add a third of this strip, which way will you go?” A child can move a finger to the right without knowing where to stop. Now uncover the answer. Two fifths ends to the left of the finger’s starting place. Something has gone wrong, and the child can be the first to notice. That leaves the teacher with a decision. I would let the child try another mark if the child can use the half to judge the last one. If the child cannot yet find the half, keeps changing which strip counts as one, or would plainly welcome a demonstration after several unsuccessful tries, I would show the missing step. Then I would hand back the pencil. I have more confidence in the question than in my rule for acting on the answer. The question reveals which part of the calculation the child can already check. The research gives a much less certain answer about which child learns more by going first. For children of this age the experiments lean toward showing first, and no trial has tested my rule. Ask where the sum must land before anyone computes. After showing the common unit, leave the next piece for the child to place. > **Figure 1.** Two fifths can be ruled out before the sum is known. The short marks divide one whole into fifths. Constructed from the example. > > What the figure shows: A line from 0 to 1 with short marks at each fifth. Halfway is marked 1/2. A dot marks 2/5, which lies to the left of halfway, so a sum that starts at a half and moves right cannot end there. The check behind that question has been printed in the Common Core standards since 2010 [1]. The decision that follows it has usually been made before the teacher sits down, by the textbook. Of three grade 5 programs examined here, two set this same sum, and the third sets 2/3 + 1/2 and says its students have met 1/2 + 1/3. One lets children try first and builds on their attempts. Another has partners say whether a half and a fourth can be added, then leads the class through folding before anyone works a sum. The third has children try, then explains. In the lessons read here, only one asks where the sum must land before it teaches the method. All three hold the same rating from the same independent reviewer, which says it rates materials “as they’re designed, not as implemented” [2, 3, 4, 5]. I write and will sell elementary mathematics books for families and schools, an interest set out below with the three programs. ## The question to which two fifths is right Suppose a child takes two shots at a basket and makes one. Later the child takes three shots and makes one. What fraction of the day’s shots went in? Two out of five. Adding the two top numbers and the two bottom numbers is exactly the right thing to do with this record. The record holds two baskets and five attempts. Now pour half a cup of water into a bowl. Add a third of the same cup. The cup has stayed on the counter, the same size all along. The water in the bowl has risen. Calling the result two fifths of a cup would ask the child to believe there is less water after the second pour. The difference is in what gets combined. In the game, two sets of attempts become a larger set, and the question asks for the share of that larger set that succeeded. In the bowl, two amounts are added and measured with the original cup. The marks on the page tell me which calculation the child did. They tell me little about whether the child was counting pieces, borrowing a rule, or writing something to get an adult to move on. I would save the basketball question for after the cup is secure. It is a good question for a teacher to have in mind before declaring that one must never add denominators. Researchers have documented the pull of whole-number arithmetic on fraction problems. They have also documented how variable an individual child’s answers can be. In a study published online in August 2026, Sabrina Di Lonardo Burr and colleagues recruited 1,055 children in 24 classrooms in Jinan, China. They assessed fraction knowledge four times across grades four and five. Understanding a shaded part of a whole tended to settle earlier. Judgments about size and about the numbers between two fractions shifted more often. Some children who answered correctly on one occasion returned to a whole-number response later. One answer can guide the next question. It is a poor basis for giving a child a mathematical identity [6, 7]. A second reason calls for care with the opening. A child may know that adding a positive length carries the finger forward and still be unable to place 2/5. That child has already shown something about addition. Offer a second copy of the whole, folded into five equal parts. Ask the child to lay two of those parts against the half. A child who now rejects the answer needed help comparing fractions. Starting the entire addition lesson again would have obscured the distinction. ## What the teacher can give away Lay three paper strips of equal length on the table. Leave one whole. Fold another in half and a third into thirds. Take one half and one third, put the half at the left end of the whole and slide the third up against it. The join and the far end are the two pencil marks that matter. Ask for a way of reporting the total to someone who cannot see the table. Try thirds as a ruler. The join falls between two marks. Try halves. The join lands, but the far end does not. Now split each third of the ruler into two equal steps. Six of the little steps fill the whole. Three reach the join and two more reach the far end. Here is what the written line records. 1/2 + 1/3 = 3/6 + 2/6 = 5/6. > **Figure 2.** On the sixths scale, three steps reach the join and five reach the far end. The rulers all measure the same whole. Constructed from the example. > > What the figure shows: A half and a third laid end to end on one whole, labeled 1/2 and 1/3. Three rulers measure the same whole. In thirds, the join and the far end both fall between marks. In halves, the join lands on a mark and the far end does not. In sixths, three steps reach the join and five reach the far end. The child has done two quite different things in that line of arithmetic. First the child has renamed two lengths without changing either. Only then has the child added the counts of steps of one size. Ask one more question. “When the half became three sixths, what got smaller?” Each step did. The length covered by the three steps is still the half put down at the start. Six is convenient. Twelve also works. The half covers six twelfths, the third covers four, and ten twelfths ends at the same pencil mark. A child who proposes twelfths has supplied a solution. Asking for the smallest common denominator at this moment can turn a mathematically successful choice into what sounds like a mistake. There will be time to discover why smaller counts are easier to manage. The whole matters even when the drawing looks convincing. Half of a twelve-centimeter strip and a third of an eighteen-centimeter strip are both six centimeters. Those lengths can certainly be added. Together they make twelve centimeters, one whole of the first strip or two thirds of the second. Before saying what fraction of a whole the result is, someone has to choose the whole. Bare numbers such as 1/2 and 1/3 already share the unit 1. Paper, ribbons and pictures require the teacher to say which length stands for it. The Common Core’s grade 3 standard says comparisons “are valid only when the two fractions refer to the same whole” [1]. ## Two good beginnings for the same lesson In one version, the teacher shows the six equal steps at once. Put the half across three sixths, write the equality under it and slide the original half on top again. Then give the child the third. “Where will this end if you start it here?” The teacher has removed the search for a unit. The child still has to preserve the length when the starting place moves and decide how many of the new units it covers. Put the strips away and a harder question remains. A classmate says that turning a half into three sixths has made it three times as long. What drawing would settle that disagreement? In the other version, the teacher puts the half and the third end to end and waits while the child tries a ruler made of thirds. The end falls between the marks. The child who folds one of those intervals in half has earned another turn with the pencil. A child who repeatedly moves the marks until both pieces appear to fit needs one thing shown immediately. The size of a step stays fixed from one end of a ruler to the other. Mark the sixths on a fresh strip, cover three with the half, and leave the third for the child to place. These versions can converge in less than a lesson. A brief invitation to point or draw is also possible in either version. It differs considerably from the extended search researchers sometimes call productive failure, in which students generate solutions to a deliberately unfamiliar problem and a teacher subsequently develops the formal idea. Calling every pause before an explanation “productive failure” would let a thirty-minute experiment take credit for a question asked in passing [8, 9]. My reason for the condition is what a child can do with a mark that turns out wrong. Say the child tries a ruler marked in thirds and the half ends between two marks. A child who can use the half to judge that mark reads the miss as news about thirds, and picks the next ruler for a reason. Without the check, the miss tells the child only that the adult is still waiting. No study found for this essay has tested this reasoning, and one finding is a reason to doubt it. Robert Siegler and Hugues Lortie-Forgues write that understanding the sizes of fractions “often is insufficient” for learning to compute with them. They describe future teachers who knew the sizes well and computed poorly. They add that understanding sizes “does seem to be an essential component” of improving arithmetic [10]. Knowing where the half lies may be needed without being enough to let a child find the sixths. With a class, I would ask everyone to put one mark where the half ends and indicate the direction in which the sum must lie. I would look at the work of children who have said nothing. If several strips show a changing whole, I would model the common unit for the room. A child who finds folding difficult can point while a partner moves the paper. Each child can then make a fresh mark after the demonstration. That second mark also lets the teacher revise a mistaken first impression. In an earlier note I posed a question about a six-year-old, a bookcase in the hall and a gap beside a window. Would the bookcase fit? A child who can already compare two widths side by side has a reason to want a length of string. The string carries one width to the place where the other is, so the two can be compared. The fraction problem has made me qualify the argument I made there. A question can create a need for a tool. Whether a child can feel that need depends on what the child can already do [11]. ## An error on the record since 1978 In 1978 the National Assessment of Educational Progress put a fraction sum to more than 20,000 eighth graders. Which whole number is closest to 12/13 + 7/8? The choices were 1, 2, 19, 21 and “I don’t know.” Each fraction is a little less than one, so the sum is a little less than two. Twenty-four percent of the students chose 2. The most common answer was 19 [6]. Nineteen is 12 + 7, the two top numbers. Twenty-one is 13 + 8, the two bottom numbers. The item is the child’s 2/5, set as a multiple-choice question 48 years ago. A review by Hugues Lortie-Forgues, Jing Tian and Robert Siegler reports the 1978 figures secondhand. The same review describes a later study of sixth and eighth graders adding and subtracting fractions, in which 22% of the answers treated tops and bottoms as separate whole numbers [6]. ## The check is already in the standard The Common Core’s grade 5 standard on adding fractions asks children to use benchmark fractions and number sense to judge an answer. Its example is to “recognize an incorrect result 2/5 + 1/2 = 3/7, by observing that 3/7 < 1/2” [1]. Arizona’s fifth-grade standards carry the same example, and give estimation a place alongside exact addition and subtraction with unlike denominators [12]. A federal fractions guide published the same year works a near twin. A student estimates that 1/2 + 1/5 lies between a half and three quarters, then adds tops and bottoms and gets 2/7. The teacher, the guide says, can point out that “this answer cannot be right because 2/7 is less than 1/2.” The guide adds that teachers can ask for an estimate before students compute. It recommends the number line for establishing fractions as numbers and recommends connecting procedures to their meanings. It rates the evidence for both recommendations as moderate [13]. These are good grounds for spending class time on the size of the answer. Neither source prescribes who should make the first fold. The standard gives no order of teaching. The guide’s own illustration of 1/2 + 1/3 is one in which “teachers can demonstrate” that sixths give a common denominator [13]. The check, and the advice to ask for an estimate first, belong to those two documents. What I add is a use for the answer. The child’s reply decides who makes the next mark. The basketball record is the other thing I would add, because it shows the teacher the question to which the child’s answer is correct. ## One sum, three textbooks The method here means renaming both fractions in a common unit and adding the counts. I am writing elementary mathematics books that use problems of this kind. I write and will sell the With Every Reason books, elementary mathematics books with a kit of materials, published by my company, KAGU, LLC. They are still being written and are not yet on sale. I will sell them to families and to schools, including through district adoptions like those described below. My own books take the try-first side: every new idea arrives as a task the child attempts before she is told the rule. They do not leave her there. The reason and the rule follow, and if she stalls, the adult shows one worked example in the same sitting. Great Minds, which publishes Eureka Math, Savvas, which publishes enVision, and Kendall Hunt and Imagine Learning, which distribute Illustrative Mathematics, sell to the same schools. So my interest pulls toward finding their lessons wanting, and toward whichever answer matches my own books. I report only what the programs’ own lessons and training handouts, copies of them posted by others, and one independent reviewer say, each cited. I make no comparison with my books. Every question I put to these lessons applies to mine. No trial has tested the order in my books either. The teacher pages of IM K–5 Math, by Illustrative Mathematics, are posted openly by its distributor Kendall Hunt, though some parts need a login. In grade 5, sums with unlike denominators arrive in Unit 6, Lesson 8. The lesson’s stated purpose is for students to add fractions with unlike denominators “in a way that makes sense to them.” One activity asks for 2/3 + 1/2. The notes tell the teacher to “Monitor for and select students” who used a number line or equivalent fractions, and to have them share. For a child with no strategy they supply a question, “How could you use one number line to show both of these fractions?” [14]. In Lesson 9 students rename fractions and “use those equivalent fractions to find sums.” Its warm-up asks whether 1/2 + 1/4 = 2/4 is true. A child can reject that by size, though the notes give its purpose as “using equivalent fractions to add and subtract” [15]. Lesson 10 explains why multiplying the two denominators always yields a common one. Its notes say students had added 1/2 + 1/3 “in a previous lesson” [16]. An estimate comes in Lesson 12, in a warm-up on 3/17 + 17/19, three lessons after the method [17]. Eureka Math is published by Great Minds. The lesson read here is Module 3, Lesson 3, in Great Minds’ 2013 text as featured on EngageNY, New York State’s curriculum site, and reproduced with annotations by Achieve the Core. It is not the 2015 edition that EdReports rated, nor the current Eureka Math², which was not examined. The teacher asks partners whether one half and one fourth can be added. Then the script reads “Lead students through the process of folding,” with paper for each child, and after that, “Draw a rectangle model for students.” The second problem is 1/3 + 1/2. Before drawing it the teacher asks, “Do our units get bigger or smaller when we create like units? Talk to your partner.” It is nearly the question I ask about the sixths. No task in the lesson asks for an estimate before computing. Its closing discussion compares finished sums with a half and with one [18]. The module’s overview, dated 2014, lists Lesson 13, in which children “use fraction benchmark numbers to assess reasonableness” of sums and differences [19]. In this text, as in Illustrative Mathematics, the estimate follows the method. Less of enVision Mathematics, published by Savvas, could be read, because its teacher’s edition is not posted publicly. What follows comes from two Savvas training handouts and from lesson passages quoted in EdReports’ review. Every lesson begins with a problem that students work, called Solve & Share [20, 21]. The handout for the previous edition says that in the next step the concepts “are made explicit through direct instruction,” as the teacher guides students “through a completed problem” [21]. In the passages EdReports quotes, Lesson 7-1, the topic’s first lesson, asks for an estimate. It offers three lengths, 1/2, 1/3 and 7/8 of a yard. Children are to pick, “Without finding the exact amount,” the two that come closest to a yard and a half. Lesson 7-3 adds 1/3 mile to 1/2 mile and asks whether 12 would also serve as the common denominator [4]. In the lessons read, enVision is the only one of the three that asks for an estimate before the method. > **Figure 3.** One sum, three openings. Each lane follows one grade 5 program through its first lessons on adding fractions with different denominators. It marks who moves first, where the method is taught and where an estimate of the sum is asked. The estimate sits after the method in two programs and before it in the third. Illustrative Mathematics and Eureka Math are drawn from publicly posted copies of their lesson pages. Eureka beyond Lesson 3 is drawn from its module overview, with Lessons 4 to 12 not examined. The enVision lane rests on Savvas training handouts and on passages quoted by EdReports. > > What the figure shows: Three grade 5 programs and their first lessons on adding fractions with different denominators. Illustrative Mathematics, Unit 6: children try first in Lesson 8, the method is taught in Lessons 8 and 9, and an estimate is asked in Lesson 12. Eureka Math, Module 3: in Lesson 3 the teacher leads the class through folding and drawing and the method is taught; Lessons 4 to 12 were not examined; an estimate is asked in Lesson 13, as the module’s overview lists it. enVision Mathematics, Topic 7: an estimate is asked in Lesson 7-1, children try first in each lesson and the teacher then explains, and the method is taught in Lesson 7-3. EdReports rates all three as meeting expectations, a rating of design. EdReports is an independent reviewer whose judgments, by one account, influence adoption decisions in many states [22]. It rates the grade 5 materials of all three programs “Meets Expectations” on each of its three gateways [2, 3, 4]. EdReports “reviews materials as they’re designed, not as implemented” [5]. A rating of design does not say which opening teaches children more, and EdReports does not claim that it does. ## Who chose the book In Arizona the answer is in statute. A school district’s governing board must “select all textbooks used in the common schools.” The meetings of any committee it sets up to review textbooks must be open to the public. A copy of each textbook under consideration must sit at the district office for public review “for a period of sixty days prior to formal selection” [23]. So while an Arizona district is choosing a new grade 5 book, a parent can open each candidate to this lesson. Charter schools are exempt from this statute. In 19 states a state-appointed board reviews textbooks first and lists those that districts may consider, and the rules vary from state to state [24]. In the rest, districts are largely left to choose for themselves. Neither kind of board has much evidence of effect to go on. Morgan Polikoff and Jake Rowley wrote in Education Next in November 2026 about the evidence for curricula. The past decade, they say, “has produced vanishingly little evidence on which curricula truly work best.” Their example is a review of 36 studies of K–5 reading curricula [22]. A parent or a board member can still ask three things of the grade 5 book. Before anyone computes, does the first lesson on unlike denominators ask where the sum must land? The standard’s own example does [1]. Who makes the first mark, and if the children do, does the teacher’s guide build the explanation on their attempts? What is asked a week later, on a sum the child has not seen? These questions are not a buying guide. My own books would have to answer every one of them, and none of the answers would show which order teaches children more. ## The closest experiment used pizza Experiments on the order of a lesson exist, though none on this sum. In a German classroom study led by Claudia Mazziotti, fifth graders had to decide whether two groups sharing pizzas received the same amount per person. It is a close cousin of the strip problem. Both require a child to see two fractions on a common scale. The pizza question asks about equivalence. The strip also asks the child to join two lengths [8]. Across nine classes in three schools, 228 children aged ten to twelve took part. Classes, not children, were randomly assigned to one of four conditions. Some tried the problem before a teacher explained it, some afterwards. Some worked alone, some in pairs. Both orders included about thirty minutes of problem solving and forty-five minutes of instruction. The instructor compared typical student solutions with the conventional one, using examples collected earlier so that the demonstrations would be comparable across classes. The children had enough prerequisite knowledge, by the researchers’ measure, to get started. One pair suggested that the birthday child should receive a whole pizza. Quite a reasonable proposal for a party. It changed the arithmetic the researchers wanted solved [8]. On the subsequent test of fraction concepts, the researchers detected no advantage for starting with the problem. Only nine classes were assigned, and the authors report that their test’s internal consistency was low. The result still belongs near the front of any honest case for letting the child try. Here were children of approximately the right age, working on fractions, offered a substantial chance to think before teaching. Being able to begin did not guarantee that beginning this way paid off on the test. The study did not use a child’s response to a half-strip question to choose the next move [8]. ## The purple pen Three studies from an overlapping group of researchers let younger children tackle a different puzzle. Write 3 + 5 = 4 + □. A child who takes the equals sign as the signal to supply a total may put 8 in the box. A child who adds every visible number gets 12. The answer is 4. Each side must name the same number. This is a useful counterpart to the fraction problem because the wrong answers have reasons an adult can recognize [25, 26, 27]. Marci DeCaro and Bethany Rittle-Johnson randomly assigned 159 children in grades two through four to individual lessons. Children who tried the equations before instruction averaged 67% on conceptual questions, compared with 60% for those taught first, across the immediate and two-week assessments. No matching procedural advantage appeared. Neither age nor measured prior knowledge reliably changed the effect of the order in this sample, which excluded high pretest scorers. That one exploratory test counts, weakly, against a broad claim that a teacher can choose the better order from what a child already knows [25]. Emily Fyfe and colleagues obtained a different result with 122 second and third graders. Their lessons provided feedback and asked children to explain correct solutions. Children taught first did better on procedures, 62% against 49%, and on questions about equation structure, 52% against 43%. The groups did similarly on the other conceptual measure, knowledge of the equals sign. Hearing an explanation early seems, in this lesson, to have given children an idea they could use in explaining the next solution. The studies differed in enough respects that no single prompt can account for the reversal with confidence [26]. Then came the purple pens. In Abbey Loehr and colleagues’ first classroom experiment, 41 second graders completed the two phases in different orders. The children who tried first scored lower on conceptual understanding, 66% against 88%. In a second experiment with 47 second graders, every child eventually received a purple pen and the original work. The child checked an answer or wrote a revision above it. This time the children who had tried first did better on subsequent problem solving, with comparable conceptual scores [27]. The experiments differed in other ways too. The pens make a memorable detail because they rendered visible a part of the lesson that is easy to omit. Once children have something better to think with, do they ever see their first answer again? On the fraction page I would leave 2/5 unerased. After the sixths have been drawn, give the child a different pencil and ask for one mark showing where the old answer would have landed. Now the reason to change the answer belongs to the child. ## What the average leaves out Tanmay Sinha and Manu Kapur combined 53 studies yielding 166 comparisons of problem solving before and after instruction. Kapur is the researcher who developed productive failure, so the review weighs one of its authors’ own ideas. Across ages and subjects, attempting first had an average advantage of 0.36 standard deviations on conceptual understanding and transfer to other problems (95% confidence interval, 0.20 to 0.51). On procedural outcomes, available in 51 comparisons, the estimate was −0.03 (interval, −0.20 to 0.15). If the test asks whether a student can carry out a taught procedure, the review found no reliable advantage for either order. Only 75 of the 166 comparisons were in mathematics. The review’s own summary adds that for second to fifth graders the effects favored instruction first. Its estimate for the randomized comparisons alone was 0.25 (interval, 0.04 to 0.47) [9]. An average also hides how the lessons inside it were built. An earlier review by Katharina Loibl, Ido Roll and Nikol Rummel finds that trying first fosters learning only in lessons with particular design features. It names two, contrasting cases and instruction built on the students’ own solutions [28]. Loibl and Rummel are also among the authors of the pizza study. The purple pen does this in small, and so does leaving 2/5 on the page. That review gives a way to read the three textbooks. The Illustrative Mathematics notes describe the kind of try-first lesson it supports, with the discussion built on strategies the teacher gathers from the children. Savvas’s 2020 handout tells the teacher to have students share their work and discuss each other’s ideas, with the publisher’s sample student work as another option [20]. The handout for the previous edition suggests displaying correct and incorrect strategies. It says teachers “identify strategies that they want students to share with the whole group.” The second step, it says, “connects the students’ thinking and solutions” to the lesson’s ideas [21]. The explanation that follows is a prepared, completed problem. Whether it is taught from the class’s own attempts cannot be seen from the handouts. ## The case for showing first The case for showing first does not depend on a child’s bad history. Paul Kirschner, John Sweller and Richard Clark set it out in 2006. A learner without the method has to search for it. Such a search, they write, “can function perfectly with no learning whatsoever.” A worked example removes the search and points attention at the moves that matter. Problem solving becomes the better use of time only “when learners are sufficiently experienced” that a worked example has become redundant. They call this the expertise reversal effect [29]. On that reading my check sits in the wrong place. It tests a prerequisite, the size of a half. The lesson teaches the common unit, and a child who can find the half has not yet met one. For this lesson that child is still a novice. For a child who has spent several lessons guessing at fractions, the next invitation to guess has a history. Showing a half laid neatly across three sixths can restore the possibility of a conversation. Both people can point to precisely what they are talking about. A teacher who knows that history has a good reason to start there. The strongest experiment on that side has nothing to do with fractions. Greg Ashman, Slava Kalyuga and John Sweller ran two fully randomized experiments with Year 5 primary students learning about the energy efficiency of light. In the first, with 64 students, those taught explicitly before solving problems did better on problems like the ones in the lesson, with no difference on transfer problems. In the second, with 71 students and harder material, they did better on both. The authors conclude that when a topic has many interacting parts, “explicit instruction should precede problem-solving” [30]. Sweller developed cognitive load theory, which predicts the result. The Sinha and Kapur review’s youngest subgroup, grades two through five, deserves particular care. The subgroup table prints an estimate of −0.09 with an interval of −0.92 to −0.16. The estimate falls outside its own interval. Four of the other ten rows of the review’s Table 4 have the same defect. Both the table and the discussion point toward showing first for this group, but those figures cannot establish a trustworthy magnitude or degree of certainty. The authors themselves write that younger students “seemed to benefit” from instruction first, though in 18 of the 25 comparisons the children had produced several solutions of their own. Nor were children randomly assigned to ages. Older and younger groups studied different material in different settings. The evidence offers no age cutoff a teacher could responsibly use. The review includes some of the studies discussed here, so its average and the individual experiments are partly the same evidence [9]. Substantial evidence sits on the other side of the teaching decision. Christina Barbieri and colleagues synthesized 55 studies, spanning elementary through postsecondary mathematics, on learning with worked examples. The average effect on mathematics performance favored examples by 0.48 standard deviations. This is a different comparison. An experiment can show that studying a solved example beats solving the same problems without one. It does not settle whether a child should get a first look at one problem before seeing that example. Comparing 0.48 with 0.36 as if the larger number won would manufacture an answer neither review estimated [31]. The strongest evidence for explicit teaching comes from children who struggle with mathematics. Madhavi Jayanthi and colleagues randomly assigned 205 fifth graders experiencing mathematics difficulties to a small-group fractions intervention or usual instruction. The intervention combined systematic teaching, rods and number lines, with a focus on explaining solutions. Reported effects favored the intervention on procedures (1.07 standard deviations) and on fraction proficiency and understanding (0.66 to 0.78). The What Works Clearinghouse rated the 2018 technical report on the same trial, which analyzed 186 students, as meeting its design standards without reservations. This was sustained teaching over months. The result gives a teacher reason to demonstrate and ask for explanations with confidence. It cannot show what proportion of the gain came from the order of the first two minutes [32, 33]. A trial published in 2026 by Robin Schumacher and colleagues makes the distinction especially clear. Ninety-nine fifth graders with mathematics difficulties were randomly assigned to usual teaching or 36 lessons over twelve weeks. Children in the intervention examined correct and incorrect worked solutions and explained what the steps meant. They outperformed the comparison group on verbal and written explanations (effects of 0.89 to 1.11 standard deviations), as well as on fraction proficiency. A child who has seen a method can still have a great deal of mathematical thinking left to do. The experiment tested the full program, and did not isolate either the worked examples or their timing [34]. A review by Sarah Krowka and colleagues found an average benefit of 0.68 standard deviations across 28 rational-number intervention studies involving U.S. students with mathematics difficulties in grades three through nine. The 95% confidence interval for that average was 0.51 to 0.85. The prediction interval, which describes the much greater uncertainty about an effect in a new setting, stretched from −0.36 to 1.8. Programs that used number lines or taught mathematical language tended to show larger gains in adjusted analyses. Those features traveled with other features of good teaching. Several review authors also wrote the fifth-grade trials or served on the federal intervention guide. Taken together, the work supports giving struggling students coherent instruction and a shared vocabulary. It supplies no timetable for how long any one child should struggle [35, 36]. Two more federal guides lean the same way. A 2007 guide recommends, on evidence it rates moderate, that students alternate between worked solutions and problems of their own [37]. The 2021 intervention guide rates the evidence for systematic instruction strong, and describes worked examples from which steps are left out for the child to supply [36]. Showing first saves time and spares a child one more failure. Its cost is that the teacher never sees what the child would have tried, and a child who could have found the sixths is handed them. My answer stands in part, and the rule at its center is not the part. The opening question stands. It takes under a minute, and the standard already prints its check. Leaving the next piece to the child stands too. If I show too soon, the child still has the third to place, and the 2021 guide’s worked example with a step left out is the same move. The rule between those two does not stand as an answer. Nothing above shows that a child who can use the half learns more by making the next mark before seeing the common unit. For children of this age the experiments lean the other way, though DeCaro’s and Loehr’s second experiment did not, and the one fifth-grade fraction study found no difference. Ashman’s, Fyfe’s and Loehr’s children were not chosen for difficulty, and many of them would have passed my check. Waiting has a price as well. In Loehr’s first class, the children who tried first and were taught afterward scored 22 points lower on concepts than those taught first. So I hold the rule as a hypothesis for a trial. Until one is run, a teacher who shows the common unit first and hands back the pencil is on firmer ground than I am when I wait. ## Check I would judge my own page by what the child can do after it closes. Here are three questions I would keep for another day, without the original picture in view. **Will two thirds plus three quarters pass one?** Each is more than half, so the sum must pass one. Now find the endpoint. Eight twelfths and nine twelfths give seventeen twelfths, five twelfths beyond one. A child who predicts the crossing and then makes an arithmetic slip has shown the teacher exactly where to look. **Will eighths do for three quarters plus one eighth?** Six eighths and one eighth make seven eighths. Thirty-seconds would also do. Ask the child to show why twenty-eight thirty-seconds ends at the same point. This time multiplying the two denominators produces much more work than the sum needs. The choice tests whether the child knows what a common unit is for. **Half a meter of ribbon is joined to a third of a meter. A friend says the length is four fifths of a meter. Could that be right?** Four fifths passes the first test. It is longer than a half and shorter than one. To settle it, compare the proposed length with five sixths. In thirtieths they are 24 and 25. The difference is just one thirtieth of a meter. The opening question was powerful enough to catch an impossible answer. Here the child discovers why an answer that looks reasonable still needs a proof. A number line can become a routine as readily as a rule can. Sevim Nuraydin and colleagues randomly assigned 188 students in grades five through eight to versions of a fraction number-line game or an active comparison. The younger students improved on the kind of zero-to-one line they had practiced. The gain did not carry to a zero-to-five line, fraction comparison, or fraction arithmetic. The game supplied feedback and omitted direct instruction [38]. Even a teacher with no intention of using a screen can take a lesson from this. Try the same question on a line with a different endpoint. A child who has learned a position on the page can put a half in the middle of a zero-to-five line. A child who checks the unit will put it between zero and one. The same caution applies to my own preference for teaching the reason. In a preregistered experiment with 156 students in grades six through nine, D. W. Braithwaite and G. J. Hall gave explicit instruction in fraction arithmetic. Varying whether students heard rationales for procedures, or whether practice alternated different procedures, produced no detectable extra gain in procedural accuracy. The students were older, and the outcome was calculation. The result has a consequence for anyone writing this kind of lesson. Printing an excellent reason is an accomplishment for the author. Finding out whether the reader can use it requires a fresh problem [39]. In deciding whether to keep this way of teaching, I would also want a test the page could fail. Before assigning classrooms to the two openings, record what children can do with a fixed whole and a benchmark. Give both groups the same worked explanation, practice time and opportunity to revise. Score a fresh problem after a delay, with the scorer unaware of the lesson order. The telling comparison would be whether the benefit of attempting first differs with the starting knowledge that was measured. If children who show the benchmark gain just as much from seeing the method immediately, the case for making them wait grows weaker. If children who initially miss it benefit from trying first, the gate is too narrow. No study found for this essay tested that particular decision rule. The evidence against me begins with Ashman’s two experiments and the results for younger children in Fyfe’s study and Loehr’s first class. Sinha and Kapur read their own youngest group the same way, and DeCaro’s test found no effect of prior knowledge. The arithmetic is certain, and the opening question is printed in a standard and a federal guide. The rule that follows it rests on reasoning from the strip and on no trial, and it is the part I hold loosely. The rule is for one child at a table, and for a class it offers only a look at every child’s first mark. The comparison of programs has gaps. The enVision teacher’s edition was not seen. Of Eureka, one lesson and the module overview were read, in the EngageNY text of 2013 and 2014. EdReports rated the 2015 edition of Eureka, and the 2021 Illustrative Mathematics in another distributor’s edition. No figure for how many classrooms use each program was found, so nothing here ranks the programs by reach. The 1980 article that first published the 1978 item was not retrieved, and no later national result on adding fractions was found. The Ashman, Loibl and Barbieri papers were read in abstract only. No publisher was asked to comment before publication, and a factual error that any of them shows will be corrected here with a dated note. What I would keep from the first moment at the table is modest. The child wrote two fifths. Before the teacher said “sixths,” there was already one fact the child could own. The finger had to move to the right. A related question arises when a school offers [eighth-grade algebra](https://barani.org/writing/what-eighth-grade-algebra-replaces/). Where will students learn the mathematics the displaced course was meant to teach? ## Keep Show the step the child needs, and leave the next one to the child. ## Sources 1. Common Core State Standards Initiative. Common Core State Standards for Mathematics, 2010. See 3.NF.3d, the footnote to the grade 4 fraction standards, and 5.NF.1 and 5.NF.2. [Standards](https://corestandards.org/wp-content/uploads/2023/09/Math_Standards1.pdf). 2. EdReports. Imagine Learning Illustrative Mathematics K–5 Math (2021). [Review](https://edreports.org/reports/overview/imagine-learning-illustrative-mathematics-k-5-math-2021). 3. EdReports. Eureka Math (2015). [Review](https://edreports.org/reports/overview/eureka-math-2015). 4. EdReports. enVision Mathematics Common Core (2024), fifth grade. Source of the ratings and of the quoted passages from Lessons 7-1 and 7-3. [Review](https://edreports.org/reports/detail/envision-mathematics-common-core-2024/fifth-grade/gateway-one). 5. EdReports. Q&A: Evidence and efficacy in curriculum adoption. Interview with Eric Hirsch, executive director, April 23, 2024. [Article](https://www.edreports.org/resources/article/qa-evidence-and-efficacy-in-curriculum-adoption). 6. Hugues Lortie-Forgues, Jing Tian and Robert S. Siegler. Why Is Learning Fraction and Decimal Arithmetic So Difficult? Developmental Review, 2015. Research review, read in the authors’ manuscript. It reports the 1978 assessment item from T. P. Carpenter and colleagues, Mathematics Teacher, 1980, which was not retrieved. [Review](https://files.eric.ed.gov/fulltext/ED565462.pdf). 7. Sabrina M. Di Lonardo Burr and colleagues. Finding their way: Students’ fraction misconceptions and errors over two years of fraction learning. Learning and Instruction 106, 102454, 2026. Published online August 6, 2026. [Four-wave longitudinal study](https://eprints.whiterose.ac.uk/id/eprint/244198/). 8. Claudia Mazziotti, Nikol Rummel, Anne Deiglmayr and Katharina Loibl. Probing boundary conditions of Productive Failure and analyzing the role of young students’ collaboration. npj Science of Learning, 2019. [Fifth-grade classroom quasi-experiment](https://www.nature.com/articles/s41539-019-0041-5). 9. Tanmay Sinha and Manu Kapur. When Problem Solving Followed by Instruction Works: Evidence for Productive Failure. Review of Educational Research, 2021. See the pooled effects and the grade-level row of Table 4. [Meta-analysis](https://doi.org/10.3102/00346543211019105). 10. Robert S. Siegler and Hugues Lortie-Forgues. Hard Lessons: Why Rational Number Arithmetic Is So Difficult for So Many People. Current Directions in Psychological Science 26(4), 346–351, 2017. Read in the accepted version. [Article](https://eprints.whiterose.ac.uk/id/eprint/120191/). 11. Igor Barani. Why should a new idea in mathematics start with a problem a child cannot yet solve? November 5, 2026. [Earlier note](https://barani.org/writing/the-need-before-the-idea/). 12. Arizona Department of Education. Arizona Mathematics Standards, 5th Grade. Adopted December 2016, updated January 2025. See 5.NF.A.1 and 5.NF.A.2. [Standards](https://www.azed.gov/sites/default/files/2025/03/Math%20Grade%205%20Final%202025.pdf). 13. Robert Siegler and colleagues. Developing Effective Fractions Instruction for Kindergarten Through 8th Grade. Institute of Education Sciences, 2010. See recommendations 2 and 3. [Practice guide](https://ies.ed.gov/ncee/wwc/Docs/PracticeGuide/fractions_pg_093010.pdf). 14. Illustrative Mathematics. IM K–5 Math, Grade 5, Unit 6, Lesson 8, Add and Subtract Fractions. Preparation and lesson pages, read November 19, 2026. [Preparation page](https://curriculum.illustrativemathematics.org/k5/teachers/grade-5/unit-6/lesson-8/preparation.html) and [lesson page](https://curriculum.illustrativemathematics.org/k5/teachers/grade-5/unit-6/lesson-8/lesson.html). 15. Illustrative Mathematics. IM K–5 Math, Grade 5, Unit 6, Lesson 9. Read November 19, 2026. [Teacher page](https://curriculum.illustrativemathematics.org/k5/teachers/grade-5/unit-6/lesson-9/lesson.html). 16. Illustrative Mathematics. IM K–5 Math, Grade 5, Unit 6, Lesson 10. Read November 19, 2026. [Teacher page](https://curriculum.illustrativemathematics.org/k5/teachers/grade-5/unit-6/lesson-10/lesson.html). 17. Illustrative Mathematics. IM K–5 Math, Grade 5, Unit 6, Lesson 12. Read November 19, 2026. [Teacher page](https://curriculum.illustrativemathematics.org/k5/teachers/grade-5/unit-6/lesson-12/lesson.html). 18. New York State Common Core Mathematics Curriculum (EngageNY), Grade 5, Module 3, Lesson 3, 2013, as reproduced with annotations by Achieve the Core, 2015. [Lesson](https://achievethecore.org/content/upload/Grade%205%20EngageNY%20lesson%20-%20add%20fractions%20with%20unlike%20units%20FINAL02.03.15.pdf). 19. New York State Common Core Mathematics Curriculum (EngageNY), Grade 5, Module 3, Module Overview, 2014, as hosted by UnboundEd. See Topic D, Lessons 13 and 15. [Overview](https://unbounded-uploads.s3.amazonaws.com/attachments/2153/math-g5-m3-module-overview.pdf). 20. Savvas Learning Company. enVision Mathematics © 2020, Teaching a Lesson. Training handout. [Handout](https://mysavvastraining.com/assets/files/documents/Handout_enVision%20Mathematics_Teaching%20a%20Lesson_1673539001.pdf). 21. Savvas Learning Company. enVisionmath2.0 © 2017, Problem-Based Learning. Training handout for the previous edition, 2020. [Handout](https://mysavvastraining.com/assets/files/documents/SAV_Handout_enVisionmath2-0_PBL_1550292728_1608674759.pdf). 22. Morgan Polikoff and Jake Rowley. The Myth of the Evidence-Based Curriculum. Education Next, November 12, 2026. Feature article. It describes a review of 36 studies of K–5 English language arts curricula. [Article](https://www.educationnext.org/wp-content/uploads/2026/11/ednext_26_4_feature_polikoff_rowley.pdf). 23. Arizona Revised Statutes §15-721, subsection F, paragraphs 1 to 3. For the exemption of charter schools, see §15-183, subsection E, paragraph 5. [Statute](https://www.azleg.gov/ars/15/00721.htm) and [charter schools](https://www.azleg.gov/ars/15/00183.htm). 24. Sy Doan and Julia Kaufman. What Role Do States Play in Selecting K-12 Textbooks? The State Education Standard 24(1), National Association of State Boards of Education, January 2024. [Article](https://www.nasbe.org/?p=23236). 25. Marci S. DeCaro and Bethany Rittle-Johnson. Exploring mathematics problems prepares children to learn from instruction. Journal of Experimental Child Psychology, 2012. [Individual tutoring experiment](https://pubmed.ncbi.nlm.nih.gov/22849809/). 26. Emily R. Fyfe, Marci S. DeCaro and Bethany Rittle-Johnson. An alternative time for telling: When conceptual instruction prior to problem solving improves mathematical knowledge. British Journal of Educational Psychology, 2014. [Individual tutoring experiment](https://pubmed.ncbi.nlm.nih.gov/24494594/). 27. Abbey M. Loehr, Emily R. Fyfe and Bethany Rittle-Johnson. Wait for it . . . Delaying instruction improves mathematics problem solving: A classroom study. The Journal of Problem Solving, 2014. [Two second-grade experiments](https://docs.lib.purdue.edu/jps/vol7/iss1/5/). 28. Katharina Loibl, Ido Roll and Nikol Rummel. Towards a Theory of When and How Problem Solving Followed by Instruction Supports Learning. Educational Psychology Review 29, 693–715, 2017. Read in abstract. [Review](https://cris.iucc.ac.il/en/publications/towards-a-theory-of-when-and-how-problem-solving-followed-by-inst/). 29. Paul A. Kirschner, John Sweller and Richard E. Clark. Why Minimal Guidance During Instruction Does Not Work: An Analysis of the Failure of Constructivist, Discovery, Problem-Based, Experiential, and Inquiry-Based Teaching. Educational Psychologist 41(2), 75–86, 2006. [Article](https://doi.org/10.1207/s15326985ep4102_1). 30. Greg Ashman, Slava Kalyuga and John Sweller. Problem-solving or Explicit Instruction: Which Should Go First When Element Interactivity Is High? Educational Psychology Review 32, 2020. Read in abstract. [Two randomized experiments](https://doi.org/10.1007/s10648-019-09500-5). 31. Christina A. Barbieri, Dana Miller-Cotto, Sarah N. Clerjuste and Kamal Chawla. A meta-analysis of the worked examples effect on mathematics performance. Educational Psychology Review, 2023. Read in abstract. [Meta-analysis](https://eric.ed.gov/?id=EJ1364058). 32. Madhavi Jayanthi and colleagues. Improving Struggling Fifth-Grade Students’ Understanding of Fractions: A Randomized Controlled Trial of an Intervention That Stresses Both Concepts and Procedures. Exceptional Children, 2021. [Randomized trial](https://journals.sagepub.com/doi/10.1177/00144029211008851). 33. What Works Clearinghouse. Review of “Impact of a Tier 2 Fractions Intervention on Fifth-Grade Students’ Fractions Achievement: A Technical Report.” Institute of Education Sciences, reviewed November 2019. [Study review](https://ies.ed.gov/ncee/WWC/Study/86431). 34. Robin F. Schumacher and colleagues. Using Worked Examples to Enhance Students’ Mathematical Explanations during Fractions Intervention. The Elementary School Journal, 2026. [Randomized trial, 36 lessons](https://doi.org/10.1086/742132). 35. Sarah K. Krowka and colleagues. Identifying key instructional practices in rational number interventions: A systematic review and meta-analysis. Learning Disability Quarterly 48(4), 2025. Published online 2024. [Meta-analysis](https://doi.org/10.1177/07319487241288823). 36. Lynn S. Fuchs and colleagues. Assisting Students Struggling with Mathematics: Intervention in the Elementary Grades (WWC 2021006). What Works Clearinghouse, Institute of Education Sciences, 2021. [Practice guide](https://ies.ed.gov/ncee/wwc/PracticeGuide/26). 37. Harold Pashler and colleagues. Organizing Instruction and Study to Improve Student Learning (NCER 2007-2004). Institute of Education Sciences, 2007. See recommendation 2. [Practice guide](https://ies.ed.gov/ncee/WWC/Docs/PracticeGuide/20072004.pdf). 38. Sevim Nuraydin, Johannes Stricker and Michael Schneider. No transfer effect of a fraction number line game on fraction understanding or fraction arithmetic: A randomized controlled trial. Journal of Experimental Child Psychology, 2022. [Randomized trial](https://pubmed.ncbi.nlm.nih.gov/35078086/). 39. D. W. Braithwaite and G. J. Hall. Explaining procedures and interleaving practice in fraction arithmetic. Learning and Instruction 90, 101854, 2024. [Preregistered experiment](https://doi.org/10.1016/j.learninstruc.2023.101854). --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/which-patients-qualify-340b/ Markdown: https://barani.org/writing/which-patients-qualify-340b/index.md Published: 2026-11-15 Modified: 2026-11-18 Form: Essay Topic: Healthcare Series: Hospital finance and care, part 4 of 4 ======================================================================== # Which patients should qualify a hospital for 340B? A hospital qualifies for 340B on its inpatients and earns on its outpatients. Igor Barani · 15 November 2026 · 49 min read Updated 18 November 2026: A federal court in Maine heard argument on the hospitals’ motion to block the revised rebate pilot. No ruling had been reported by the afternoon of November 18, and the pilot is still set to begin on January 1, 2027. I will update this essay when the court rules. Healthcare · Essay · Hospital finance and care, part 4 of 4 · https://barani.org/writing/which-patients-qualify-340b/ Today the law says the patients in its beds. A hospital enters 340B on a count of inpatient days, and the money it then makes depends on outpatients the count never looks at. So an inpatient measure cannot, by itself, show that a growing outpatient drug business produces a safety-net benefit to match. Congress should require that relationship to be measured before it decides how eligibility should change. Take one hospital. Richmond Community Hospital is a Bon Secours hospital in Richmond, Virginia, in a neighborhood the New York Times described as predominantly Black and low-income [1]. Bon Secours says Medicaid patients account for 53% of the hospital’s inpatient admissions [2]. The hospital takes part in 340B, a federal program that obliges drug makers to sell certain providers outpatient drugs at a discount [2, 3]. Attached to it is a network of outpatient infusion centers, specialty clinics and pharmacies, and that network is where its 340B money is made. From September 2018 to September 2023 the hospital reported $276.5 million in benefit from the program [2]. The majority staff of the Senate health committee found that all of the savings on drugs that physicians administer came from the provider-based outpatient infusion centers, and none from the hospital itself. Bon Secours’s own letter says “substantially all” [2]. The rest was pharmacy revenue, and Figure 1 gives the amounts. Bon Secours gave the committee its reason. Opening infusion centers “in a community with a more favorable payor mix,” it wrote, let the hospital stay in its own community. The savings, it said, “have allowed RCH to remain open, despite operating at a substantial loss for decades” [2]. Its letter says the services at the hospital are provided “at a loss of more than $33 million annually” [2]. The hospital’s list of child sites, it wrote, includes infusion centers, three oncology clinics, a gynecologic oncology clinic, a rheumatology clinic and a hepatology clinic [2]. A 2022 Times investigation had described such sites as lying in wealthier parts of the city, by a trade publication’s summary [1]. Bon Secours told the committee that the Times’s account of the infusion centers was “in every way inaccurate” [2]. It says it complies with the program, and it told the Times it had provided nearly $18 million in free care to poor patients at the hospital since 2018 [1]. The record used here holds no finding against the hospital. > **Figure 1.** Richmond Community Hospital’s reported 340B benefit, September 2018 to September 2023, by where it arose. Of $276.5 million, $232.1 million was reported savings on physician-administered drugs, which came from provider-based outpatient infusion centers. Another $44.4 million was reported revenue from drugs dispensed at pharmacies, and the staff report puts 84% of that revenue at Bon Secours’s own pharmacies. Savings and revenue are different measures, and the total is their sum. Bon Secours’s first letter gave about $39.4 million for the pharmacies and 84% of their claims, over the same five years. It later told the committee that the $39.4 million left out 2018 and 2019. Figures as the Senate committee’s majority staff states them in its April 2025 report. > > What the figure shows: One bar from zero for the 340B benefit that Richmond Community Hospital reported from September 2018 to September 2023, 276.5 million dollars in all. Of it, 232.1 million dollars was reported savings on physician-administered drugs, which arose at provider-based outpatient infusion centers. The other 44.4 million dollars was reported revenue from drugs dispensed at pharmacies. By the Senate staff report’s count, 84 percent of that revenue came through Bon Secours’s own pharmacies and 16 percent through outside pharmacies under contract. The case drew the attention of a Virginia senator at once. “I’m just worried the story is going to cause a lot of people to have really serious questions about a program that our safety net providers need,” Senator Tim Kaine said [4]. He went on, “so Bon Secours needs to have an answer, and we are going to dig into it from the health committee standpoint” [4]. The committee’s majority staff later examined the hospital, and its 2025 report is the source of the figures above. None of this breaks a rule. A disproportionate share hospital qualifies for 340B by counting the Medicaid patients, and the Medicare patients on federal income support, who lie in its beds. An uninsured patient in one of those beds who is not eligible for Medicaid lowers its score. The hospital then earns the money on drugs given in its outpatient clinics and dispensed at pharmacies. Some of those outpatients are the same poor patients, and many may not be. The gate does not measure them either way. A hospital advocate can accept every word of the Richmond account and call it the program succeeding. Money made on insured patients across town kept a hospital open in a poor neighborhood. That may be what happened. The trouble is that no rule asks anyone to show it, in Richmond or anywhere else. My conclusion is that 340B’s central accountability problem is the gap between qualifying for a financial benefit and demonstrating its public return. The program can sustain valuable services. It can also reward ownership, referral capture, and control of pharmacy transactions. These functions coexist. Eliminating a hospital’s margin does not automatically help patients if the money moves to a manufacturer, pharmacy intermediary, or payer. I am sure of the diagnosis, because it is arithmetic. I am much less sure of the fix. The formula I suggest later is one option, and I can already see how it could go wrong. I lead a radiation oncology department in a hospital, and I also take part in cancer and oncology leadership for the health system I work in. That system takes part in 340B. The opinions in this essay are my own and reflect my point of view, not that of the institutions I work for. Richmond is one hospital, and its $276.5 million is a benefit, not a profit. The American Hospital Association, the AHA, says that “nearly 44% of 340B hospitals operated on a negative margin” in fiscal 2023 [5]. The evidence warrants scrutiny of the mechanism, not a presumption that every profitable prescription is improper. ## Why Congress created an institutional subsidy The program began as a repair. Medicaid’s drug rebate program, established in 1990, entitled Medicaid to a manufacturer’s best price. A deep discount to a clinic for the poor could therefore raise the rebate the manufacturer owed on every Medicaid sale. The 1992 House committee report described clinics and public hospitals losing their discounts, and it conceded that this was only one of many factors in pricing [6]. Congress answered in the Veterans Health Care Act of 1992, which added section 340B to the Public Health Service Act [3]. A manufacturer whose drugs Medicaid covers must offer covered outpatient drugs to listed providers, called covered entities, at or below a ceiling price [3]. The ceiling starts from the drug’s average manufacturer price, or AMP, in the previous quarter. AMP is a statutory average of what wholesalers pay for drugs bound for retail pharmacies and what retail pharmacies pay when they buy direct [3, 7]. From it the formula subtracts the rebate the manufacturer would owe Medicaid on the same unit [8]. For most brand-name drugs that rebate is at least 23.1% of AMP, and more when the drug’s price has outrun inflation [8]. So 340B is not a fixed percentage discount from list price. Covered entities include federally qualified health centers, Ryan White HIV clinics, other clinics that live on federal grants, and several kinds of hospital [3]. The committee wrote that it intended “to enable these entities to stretch scarce Federal resources as far as possible, reaching more eligible patients and providing more comprehensive services” [6]. The mechanism was institutional cross-subsidy. Lower the organization’s drug costs, and it can do more with its resources. It was not designed solely as a discount card handed to uninsured patients at a pharmacy counter. An insured patient’s prescription generating revenue for services used by someone else can therefore be consistent with the original rationale. The hard question is whether that second step happens, how much reaches it, and how it can be verified. The program’s history is narrower than that reading. A 2026 history of the law was built from 175 internal documents and 19 interviews with 18 informants, and funded by the Commonwealth Fund. It notes that the “stretch” language “does not appear anywhere in the statute” [9]. Its authors report that the drafters had in mind providers that depended on government grants and had a mission to give free and discounted drugs [9]. What hospitals now do with commercially insured patients is lawful. By this account, the drafters never discussed it. Nor was every later expansion an administrative accident. Congress deliberately added categories of eligible hospitals, including through the Affordable Care Act [3, 10]. The more defensible criticism is that the program’s commercial reach expanded faster than the ability to measure its resulting public benefit. When a prescription qualifies for a safety-net discount, what additional benefit is the patient, or the community, guaranteed to receive? For much of hospital-based 340B, the answer is less specific than either side of the political argument suggests. The Congressional Budget Office, CBO, reports that the statute “does not set requirements for how participating facilities use 340B net revenues” [10]. Ryan White clinics are an exception from outside the statute. Their grant rules treat 340B revenue as program income and list what it may pay for [11]. So hospital findings cannot be generalized to every clinic. A hospital collects more for a drug than it paid, and the gap between what the insurer and patient pay and the 340B price is the spread. The spread is not profit, because dispensing, nursing, fees and unpaid care come out of it. The [companion essay](https://barani.org/writing/budget-neutral-340b/) follows one dose through that arithmetic under Medicare’s 2027 payment rule. Three separate claims hide in any statement about the spread: money was generated, a particular organization retained it, and patients benefited because of it. Evidence for one does not establish the other two. Much of the 340B debate consists of answering the first question and presenting the answer as if it settled the third. The question before all three is which patients qualified the hospital. The sums are large. Covered entities bought $100.0 billion of drugs at 340B prices in 2025, by the count of the Health Resources and Services Administration, HRSA, the agency that runs the program. Hospitals that qualify through the test described next bought $79.2 billion of that [12]. The figures measure acquisition spending, not $100 billion of savings, profit, government outlays, or charity care. ## The gate counts inpatient days Most of the money comes through one category. A disproportionate share hospital, the category that bought the $79.2 billion, must pass three tests. It must be public, or a nonprofit with a government contract to serve low-income people. It must not buy outpatient drugs through a group purchasing organization. And on its latest Medicare cost report it must show an adjustment percentage “greater than 11.75 percent” [3, 13]. For-profit hospitals cannot join [13]. That percentage is a statutory calculation, not a finding that 11.75% of its patients are uninsured or receive charity care. It comes from Medicare’s inpatient payment system, and it is built in two steps. The first step adds two ratios of days spent in the hospital’s beds. One takes the days of Medicare patients who also receive Supplemental Security Income, or SSI, the federal cash benefit for the aged and disabled poor, and divides by all Medicare days. The other takes the days of patients eligible for Medicaid but not Medicare and divides by all patient days. Medicare’s own statement of it is “DSH Patient Percent = (Medicare SSI Days / Total Medicare Days) + (Medicaid, Non-Medicare Days / Total Patient Days)” [14]. Only days in acute-care units count, and beds used for outpatient observation are left out [15]. The two ratios have different denominators, so their sum is not the share of any group of patients. It is an index. No outpatient visit, infusion or prescription enters it. Medicare’s advisory commission, MedPAC, says so directly: “Outpatient sites affiliated with a hospital do not affect the hospital’s DSH adjustment percentage” [16]. A hospital’s director of finance told a House subcommittee the same in 2017: “The DSH percentage is a representation of inpatient days” [17]. An uninsured patient who is not eligible for Medicaid counts against the hospital. Her day enters only the denominator of the Medicaid ratio, so treating more such patients lowers the hospital’s score. The ratio turns on eligibility, not on who paid. An uninsured patient who was eligible for Medicaid that day counts in the numerator, if the hospital can prove it, and so do some patients covered through a state’s federal waiver. Patients paid for from a waiver’s pool for uncompensated care do not [15]. An example shows the size of it, and its numbers are invented for the purpose. A hospital has 100,000 inpatient days a year. Medicare patients account for 40,000 of them, and 3,200 of those belong to patients on SSI, so the first ratio is 8%. Medicaid patients without Medicare account for 20,000 days, so the second ratio is 20%. The index is 28%. Now the hospital admits more poor patients, enough to add 5,000 days. If they are on Medicaid, the second ratio becomes 25,000 ÷ 105,000, or 23.8%, and the index rises to 31.8%. If they are uninsured and not eligible for Medicaid, the ratio becomes 20,000 ÷ 105,000, or 19.0%, and the index falls to 27.0%. The patients are equally poor in both cases. Only their eligibility for Medicaid differs. > **Figure 2.** The index behind the 340B gate, for an example hospital, as two ratios of inpatient days. Adding 5,000 days of Medicaid patients raises the index from 28.0% to 31.8%. The same days from uninsured patients who are not eligible for Medicaid lower it to 27.0%, because such a day enters only a denominator. Example numbers, computed with the definitions in 42 C.F.R. § 412.106. > > What the figure shows: Four strips of inpatient days for an example hospital. Of 40,000 Medicare days, 3,200 belong to patients on Supplemental Security Income, a ratio of 8.0 percent. Of 100,000 patient days, 20,000 are Medicaid days, a ratio of 20.0 percent, so the index is 28.0 percent. Adding 5,000 Medicaid days, which count in the numerator and the denominator, raises the index to 31.8 percent. Adding 5,000 days of uninsured patients who are not eligible for Medicaid, which count in the denominator only, lowers it to 27.0 percent. The second step turns the index into the adjustment percentage. For an urban hospital with 100 or more beds, the regulation sets the adjustment at 5.88% plus 82.5% of the amount by which the index exceeds 20.2% [15]. Run that backward. An adjustment of 11.75% is 5.87 points above 5.88%, and 5.87 ÷ 0.825 is 7.1, so the index must exceed 20.2 + 7.1, or about 27.3%. The line that the statute writes as 11.75% is therefore a sum of two inpatient ratios above about 27.3%. Return to the example hospital. At an index of 28% its adjustment is 5.88 + 0.825 × 7.8, or 12.3%, and it qualifies. With the uninsured patients its index is 27.0%, its adjustment is 11.5%, and it does not. With the Medicaid patients its adjustment is 15.5%. A hospital at the line that admits more uninsured patients, none of them eligible for Medicaid, can lose the discount for doing it. > **Figure 3.** How Medicare’s formula turns the index of inpatient days into the adjustment percentage, for an urban hospital with 100 or more beds. The statute’s line of 11.75% meets the formula at an index of about 27.3%, and the example hospital sits just above it at 28.0%. With 5,000 more days of uninsured patients who are not eligible for Medicaid, it sits just below, at 27.0%. Drawn from the formula in 42 C.F.R. § 412.106(d)(2)(i). > > What the figure shows: A rising straight line shows Medicare’s formula for an urban hospital with 100 or more beds: the adjustment percentage is 5.88 percent plus 82.5 percent of the amount by which the index of inpatient days exceeds 20.2 percent. A dotted level line marks the statute’s 11.75 percent, and it meets the formula at an index of about 27.3 percent. The example hospital, at an index of 28.0 percent, has an adjustment of 12.3 percent and qualifies. With the Medicaid days added, at 31.8 percent, its adjustment is 15.5 percent. With the days of uninsured patients who are not eligible for Medicaid added, at 27.0 percent, its adjustment is 11.5 percent, below the line, and it does not qualify. Other hospitals qualify by other tests, and none of them looks at outpatients. Critical access hospitals, the small rural hospitals Medicare pays at cost, face no such percentage at all. Rural referral centers and sole community hospitals need 8% [3]. Small urban and most rural hospitals have their adjustment capped at 12%, so for them the line sits just under the ceiling of the scale [15]. ## Child sites qualify through the parent’s cost report Once the parent hospital passes, its outpatient sites follow by an accounting test. HRSA asks whether an off-site clinic is listed as reimbursable on the hospital’s latest Medicare cost report, with outpatient costs and charges of its own [18]. A clinic that meets the test registers as a child site. HRSA’s test does not ask whom the site serves. A hearing record shows the sequence at one hospital. Northside Hospital’s director of finance told a House subcommittee in October 2017 that the hospital’s talks to acquire oncology practices began in 2011 and were completed in 2012 [17]. The locations operating as hospital outpatient departments were child sites, she said, registered in April 2014. Her explanation was that a large oncology practice had approached Northside “seeking integration with the hospital system, as were several other hospital systems in the Atlanta area,” for “clinically-integrated care” [17]. Asked whether patients had been denied care since, she answered, “Since we acquired them, no” [17]. The record establishes a sequence: acquisition, outpatient-department operation, registration, and hospital billing. It does not show why the hospital acquired the practices beyond its stated reason, whether any patient’s bill changed, or what the clinical effect was. Nor does the program ask about the patient’s means. HRSA’s answer to whether patients outside Medicaid can receive 340B drugs is “Yes, as long as they are patients of the covered entity” [19]. So a patient’s commercial insurance or relatively high income does not by itself disqualify the prescription. The limit is the care relationship. HRSA’s 1996 definition excludes a person “if the only health care service received by the individual from the covered entity is the dispensing of a drug or drugs” [20]. Contract pharmacies widened the reach further. A contract pharmacy is an outside pharmacy that dispenses on a covered entity’s behalf. In 2010 HRSA allowed each entity an unlimited number of contract pharmacies, where its earlier guidance had allowed one [21]. They grew from about 1,300 that year to nearly 20,000 in 2017 [21]. The pharmacy need not keep a separate shelf. In the common arrangement it fills the prescription from its ordinary stock at the ordinary price. Software later combs the filled prescriptions and patient records for eligible ones, and the entity orders 340B-priced drugs to replenish the stock [22]. That description comes from a 2026 federal court order that relied on a drug maker’s declaration. So the program’s growth is not the same as more medical treatment. CBO found that purchases through the program’s prime vendor rose from $6.6 billion in 2010 to $43.9 billion in 2021, in 2021 dollars. It judged the integration of hospitals with off-site clinics the largest of three program-specific causes, and it could not put numbers on them [10]. In 2021, 87% of that spending was on drugs administered or distributed in hospital outpatient departments and their off-site clinics [10]. A trade-press account of CBO’s figures puts off-site clinics in the program at about 6,100 in 2013 and 27,700 in 2021 [23]. An existing treatment can enter the 340B accounts when ownership, site registration, or prescription attribution changes. One study of that growth comes with disclosures worth reading. A 2025 analysis covered more than 28,000 drug codes. It attributed an average of 79.6% of 2018–2024 growth at list prices to utilization, meaning more units, and close to all of it at 340B prices [24]. The National Pharmaceutical Council, which describes itself as supported by leading biopharmaceutical companies, partly supported the study and reviewed the manuscript. IQVIA, which employs two of the three authors and owns the data, kept the final say on the text, and the data are not public [24, 25]. One of the three authors disclosed that he was an independent director of Kalderos, which, the disclosure says, “offers 340B solutions” [24]. The appointment itself can carry the value, and some companies say so in their marketing. In 2021 the consultant Advis pitched hospitals a medication-management clinic, to which affiliated physician offices refer patients for a qualifying visit. The clinic, Advis wrote, could add “340B prescription revenue from patients that wouldn’t otherwise qualify for the Program” [26]. Another company, CPS, markets “referral capture” to “retain escaping patients” [27]. Both describe practice that the statute allows, as a federal court read it. That court struck down HRSA’s narrower patient definition as “contrary to the plain language” of the law [28]. The hospital does not need the patient to become poorer, the medicine to become more effective, or the manufacturer to lower its ordinary price for the opportunity to change. An extra appointment might uncover an interaction, improve adherence, or coordinate care. It might also impose travel, time away from work, and another bill. Calling every qualifying visit a sham is unsupported. Treating a qualifying visit as evidence of clinical value is equally unsupported. ## Payment rules decide where the margin is made The spread on a prescription depends on who pays for it, and federal rules decide much of that. Medicaid’s fee-for-service program, in which the state pays the claim itself, pays close to cost. Federal guidance since 2016 has told states to pay pharmacies their actual acquisition cost plus a dispensing fee. For 340B drugs, it says, “reimbursement should not exceed the 340B ceiling price” [29]. On a Medicaid fee-for-service prescription the spread is close to nothing, by rule. HRSA also keeps those patients out of contract pharmacies by default. A drug maker must not pay twice on one unit, once as a 340B discount and again as a Medicaid rebate. To prevent it, HRSA bars 340B drugs for Medicaid fee-for-service patients at contract pharmacies unless the pharmacy, the entity and the state agree on a method [30, 31]. The inspector general found in 2014 that most entities it studied, a purposive sample of 30, did not dispense 340B drugs to Medicaid patients there [32]. Both rules are written for fee-for-service. Medicaid managed care, in which a private plan pays the claim, falls outside them, so Medicaid is not absent from the spread. An uninsured patient pays what she can, which is often little. In GAO’s sample of 55 entities, 30 reported giving low-income uninsured patients a discount at some or all of their contract pharmacies [21]. So the gate does not measure the outpatient population on which the financial opportunity depends, and the payment rules pull the two apart. Medicaid’s fee-for-service program pays close to the 340B price by rule, and its patients are kept out of contract pharmacies by default. The patients the gate counts most, Medicaid patients, therefore produce less of the margin. MedPAC writes that the formula “is primarily driven by Medicaid patient shares” [33]. The same person can still stand on both sides. A patient in a Medicaid managed-care plan counts toward the gate when she is admitted and can produce a spread when she fills a prescription. Minnesota has published 340B revenue by payer. Its health department estimates that in 2024 commercial insurers produced about 45% of covered entities’ net 340B revenue, and Medicare about 36%. The state’s public programs, Medical Assistance and MinnesotaCare, produced about 19% [34]. The figures are estimated, self-reported and unaudited, and they leave out self-pay and uninsured patients. They are also a share of revenue, a different kind of number from the gate’s index of days. They are set beside the index only to show direction. Maine now requires 340B hospitals to report by payer type as well, beginning July 1, 2026. Washington requires it where the data are available, under a law in force since June 11, 2026 [35, 36]. Neither state’s payer figures are used here. A study of cancer drugs shows the same dependence on the payer. Among Blue Cross Blue Shield patients at 1,541 hospitals from 2020 to 2024, the estimated prices hospitals paid for three biosimilars fell by 60%, 72% and 63%. What insurers paid fell by 32%, 36% and 34% [37]. The gap between them, which the authors call the markup margin, grew as the drugs got cheaper. The acquisition prices were estimated from Medicare’s price files and never observed. The study’s funder is a foundation whose board includes representatives of Blue Cross plans, the payers of the markups it measures [37, 38]. The narrower lesson is valuable: lower-priced treatment and higher provider margin can occur together if acquisition prices fall faster than reimbursement. This is also why “manufacturer-funded” does not mean “without consequences for taxpayers.” The mandated acquisition discount comes from the manufacturer. The reimbursement that turns it into a margin can come from Medicare, an employer-sponsored plan, and the patient. ## 11.75% was chosen to qualify particular hospitals The line has a recorded origin, set out in the same 2026 history [9]. One staffer recalled lowering it to 12.5% so the University of Virginia’s hospital would qualify. Senator Hatch’s condition was that the University of Utah’s hospital qualify too. So, “to accommodate Senator Hatch,” staff “found the new necessary minimum eligibility criterion: 11.75%” [9]. The change added 91 hospitals and appeared only in the enrolled bill [9]. That account rests on papers and memories from 1992. It shows how the number was chosen, not whether it is a good one. Medicare has since moved its own money off this measure. MedPAC reported in 2015, citing its 2007 report, that the formula’s payments tracked uncompensated care poorly. The hospitals in the top tenth by uncompensated care provided 41% of it and received 10% of the payments [16]. Since the Affordable Care Act, 75% of what Medicare would have paid in disproportionate share money goes into a separate pool for uncompensated care, reduced as the uninsured rate falls [14]. MedPAC then designed a replacement index for Medicare’s own payments, and voted in 2023 to recommend it [33, 39]. The index has three parts. One is the share of a hospital’s Medicare claims, inpatient and outpatient, that are for low-income beneficiaries. The others are its uncompensated care as a share of revenue and, at half weight, Medicare’s share of its inpatient days [33]. Its executive director said of 340B, by a trade publication’s account, “We are not touching it” [39]. So Medicare now pays only a quarter of its disproportionate share money on the old index, while 340B still uses it for the whole decision, on $79.2 billion of purchases in 2025. ## Hospitals near the line changed The best evidence compares hospitals just above and just below 11.75%. Desai and McWilliams did so in Medicare claims. Eligibility was associated with 2.3 more hematologist-oncologists practicing in hospital-owned facilities, 230% more than expected [40]. It was also associated with lower proportions of low-income patients in hematology-oncology, and with no more safety-net care [40]. The design describes hospitals near the line, not every 340B hospital. The AHA replied that such hospitals are a “sliver” of the program and that the study used traditional Medicare data only [41]. Hospitals also cluster at the line. In 2014–2016 the number of hospitals just above 11.75% rose 41% relative to those just below. Investor-owned hospitals, which cannot use the program, showed no such change [42]. The authors list better record keeping, admitting more Medicaid patients and misreporting as possible causes. They put the share of hospitals that adjusted at “at most roughly 8%” [42]. A 2026 study followed 122 hospitals that joined 340B in 2019 against 779 that did not, in traditional Medicare claims. Joining was followed by more total claims and more drug-administration claims at newly opened off-campus departments. The rise in drug claims themselves was not statistically significant, and the volumes were small [43]. Several studies look at where new sites register. Clinics registered in 2004 or later served communities that were wealthier and better insured than those served by earlier registrants [44]. Between 2010 and 2022 the neighborhoods of newly registered child sites and contract pharmacies became less deprived, while those of newly registered hospitals did not change [45]. Another study, funded by Johnson & Johnson, found that child sites’ ZIP-code areas had a median income 28% above their parents’ [46]. These are studies of where new sites register, by neighborhood, not of whom they treat. The hospitals’ association replied to the first that its authors had not checked whether the clinics opened before or after their hospitals joined [47]. None of the studies cited here compares a parent hospital’s patients with its own child sites’ patients. What exists is one documented hospital and neighborhood data. ## Who benefits along the entire pipeline Where the whole spread can be counted, the hospitals and clinics keep most of it. In Minnesota they reported $3.04 billion in payments for 340B drugs in 2024 and $1.53 billion to acquire them, a spread of $1.51 billion [34]. They reported about $165 million in operating costs. About $137 million of that went to outside firms, $120 million to contract pharmacies and $17 million to other outside parties such as administrators, which is about 9% of the spread. The other $28 million or so was the entities’ own internal cost [34]. They kept at least $1.34 billion in net 340B revenue, and large hospitals accounted for more than 80% of it [34]. These are amounts the entities reported paying, so they are not a full measure of what intermediaries earn. The figures are self-reported and unaudited. The state calls the reimbursements for administered drugs likely inaccurate and the reporting of internal costs inconsistent, and it does not track how the net revenue was spent [34]. > **Figure 4.** What Minnesota’s covered entities reported for 340B drugs in 2024. Of $3.04 billion paid, $1.53 billion bought the drugs and about $165 million went to operating costs, about $137 million of it paid to contract pharmacies and other outside firms. At least $1.34 billion stayed as net revenue, about 89% of the $1.51 billion spread. Self-reported and unaudited; Minnesota Department of Health, report of February 2026 on 2024 data. > > What the figure shows: One bar from zero for what Minnesota’s covered entities reported for 340B drugs in 2024, 3.04 billion dollars in payments. Of it, 1.53 billion dollars bought the drugs. The rest, a spread of 1.51 billion dollars, divides into about 165 million dollars of operating costs, which include payments to contract pharmacies and administrators, and at least 1.34 billion dollars of net revenue kept by the covered entities. The figures are self-reported and unaudited. Under current law a hospital may spend that money as it chooses. Cleveland Clinic told the committee that “there is no dollar-for-dollar ‘pass on’ requirement to patients under the 340B statute,” which is a correct statement of the law [2]. Its 340B benefit was $933.7 million from April 2020 to June 2023. Of the $395.4 million in savings on physician-administered drugs, 44% came from pharmacies it owns at its child sites [2]. It told the committee that the money reaches its income statement “without being independently segregated, distributed or allocated” [2]. The 2025 staff report found that its financial-assistance policy did not cover dispensed drugs. Cleveland Clinic says it provides “extensive pharmacy-related benefits at minimal to no additional cost to patients or payers” [2]. Bon Secours told the committee that “revenue is revenue, and all revenue is used to pay for expenses incurred in pursuit of our mission” [2]. Between the drug maker and the hospital sits a layer of companies paid on each 340B transaction. A wholesaler settles the discount. A prime vendor collects a fee from suppliers on purchases under its contract. An administrator matches a prescription to a patient, and a pharmacy dispenses it [2, 48, 49]. They supply real services: dispensing, patient counseling, inventory, claims matching, and compliance support. Few of them publish what 340B earns them. Apexus has been the sole 340B prime vendor since 2004, under an HRSA contract. About nine in ten covered entities buy through it, by Senator Cassidy’s account [50]. The government does not pay it. Drug makers and distributors pay Apexus a small percentage of sales, the New York Times reported [49]. Senator Cassidy’s letter says the company generated $227 million in 2022 [50]. The Times, which the letter cites, reported that Apexus was on track for that figure [49]. The company is private and its accounts are not public [49]. Apexus says the program “is managed by” HRSA, “not by Apexus,” and that it “has no role in setting drug prices for patients or reimbursement for covered entities” [51]. An inquiry identifies questions to investigate, not established wrongdoing. Three wholesalers handle nearly all drug distribution, by PhRMA’s count, and in a 340B sale the wholesaler settles the discount by charging the difference back to the manufacturer [48]. PhRMA, the drug makers’ trade group, concedes the larger point in its own report. “While the covered entity retains the largest share of the revenue, the wholesaler and the pharmacy also benefit financially from the 340B transaction,” it says [48]. Cencora tells its investors that a rebate model “may also supplant 340B chargeback mechanisms that we administer, which could adversely affect our business” [52]. McKesson owns Macro Helix, which by McKesson’s account serves about 40% of 340B hospitals [53]. In 2025 McKesson also bought control of the business-services company of Florida Cancer Specialists, a practice that “remains independently owned” [54]. Five companies, CVS, Walgreens, Express Scripts, OptumRx and Walmart, held three-quarters of 340B contract-pharmacy relationships, according to the Senate staff report, citing Drug Channels [2]. CVS required hospitals and clinics to use its own administrator, Wellpartner, for the 340B drugs dispensed at CVS pharmacies. It told the committee that this “integrated approach” reduces “compliance problems,” and that hospitals remain free to use other pharmacies and administrators [2]. In the contracts the committee examined, Wellpartner’s fee was the greater of $4 or 10% of the spread, rising to 14% in later contract years [2]. A fee of that kind rises with the spread, which is what is paid for the drug, by insurer and patient together, less the 340B price. It does not follow the drug’s price as such. By CVS’s own count, Wellpartner’s fees were $147 million in 2019, peaked at $386 million in 2021 and were $382 million in 2023, $1.6 billion over five years [2]. That is revenue, not profit. CVS’s pharmacies also earn a dispensing fee, which CVS says compensates them for forgoing “payments from insurers” [2]. CVS tells its shareholders that the arrangement matters to it. Fewer contract-pharmacy arrangements, less use of its administrative services, or less participation by drug makers, it writes, “could materially and adversely affect the Company” [55]. Walgreens charged administrative fees of 8%, 13% and 20% in contracts the committee examined, on top of the dispensing fee, and did not report what its administrator earns [2]. Walgreens says hospitals need not use its administrator and that its fees are fair market value [2]. The committee’s report is the work of one party’s staff, and it is not an adjudicated finding. A publicly filed contract shows how a percentage fee behaves. A 2021 agreement between a Florida AIDS service organization and a specialty pharmacy sets the pharmacy’s fee on insured claims at $17 plus 12% of total reimbursement [56]. Applying that formula to a hypothetical $1,000 reimbursement produces $137 in pharmacy compensation. If acquisition cost were $600, the spread would be $400, and $263 would remain before administrator fees and other entity costs. Wellpartner’s formula, from a different arrangement, would take $40 of the same $400 spread in a first contract year and $56 at its later rate of 14%. The $1,000 and $600 are illustrative, and the fee formulas are documented. These are one historical contract and one company’s terms, not a national average or a finding of overcharging. > **Figure 5.** Two documented fee formulas from different arrangements, each drawn against the amount it is charged on. A filed pharmacy contract’s fee reaches $137 on a $1,000 payment, and CVS’s administrator fee reaches $40 on a $400 spread at its first-year rate. Each grows with its own base, the payment for the first and the spread for the second. Example amounts; formulas from an SEC-filed contract and the Senate staff report. > > What the figure shows: Two dotted rising lines on the same scales, for two fee formulas from different arrangements. Under a filed pharmacy contract the fee is 17 dollars plus 12 percent of the payment, so it rises from 17 dollars to 137 dollars on a payment of 1,000 dollars. Under CVS’s administrator formula in the first contract year the fee is the greater of 4 dollars or 10 percent of the spread, so it stays at 4 dollars up to a spread of 40 dollars and rises to 40 dollars on a spread of 400 dollars. Each fee grows with its own base, the payment for the first and the spread for the second. Some of these companies belong to larger groups. The three largest pharmacy benefit managers, PBMs, are vertically integrated with large insurers, the Federal Trade Commission’s staff found [57]. Their groups also own pharmacies that act as 340B contract pharmacies, and two of them, CVS and Cigna, own administrators that match the claims [2, 58]. So one parent company can pay for a 340B prescription through its insurer or PBM, earn a fee on it at its pharmacy, and earn another as the administrator. Whether any group does is not shown. CVS told the committee that it keeps firewalls between its administrator, its PBM and its insurer [2]. A gain in the pharmacy subsidiary is not automatically a saving for the employer, and nothing here shows what any group nets. None of these sources has any of these companies directing the program. They draw less attention than hospitals and drug makers, and they say less about 340B in public. The legal recipient of the discount and the economic recipient of its value are not necessarily the same. Services, fees, negotiated prices, and ownership determine how that value is divided. ## The lobbying is a fight over specific financial rights The public record shows organized campaigns over the discount, the patient definition, contract pharmacies, and information. It does not reveal every meeting, establish that a senator adopted a position in exchange for money, or allocate every lobbying dollar to 340B. Lobbying disclosure supplies scale, with an important limitation. PhRMA reported $7.04 million in federal lobbying for the third quarter of 2026, across every issue it works on. Its filing lists the SUSTAIN, ACCESS and SECURE bills and Senator Cassidy’s draft among them [59]. For the same quarter the AHA reported $3.84 million across all issues, and 340B Health, the association of participating hospitals, reported $330,000 across all issues [60, 61]. None of these totals measures spending on 340B. The hospitals ask to keep the gate where it is. The AHA’s letter went to Senator Bill Cassidy, the Louisiana Republican who chairs the Senate health committee. A site should qualify, the AHA wrote, if it “is part of the 340B hospital.” It should also meet Medicare’s provider-based standards and furnish “meaningful outpatient services to the hospital’s patients” [62]. In October the AHA asked the sponsors of the SUSTAIN bill to “Remove Child Site Restrictions,” including a three-year wait for newly acquired sites [63]. On patients, the AHA’s answer to Senator Cassidy’s draft was direct. It “respectfully reject[s] the premise that the current structure of the 340B program does not ensure direct patient benefit” [62]. Every dollar saved, it wrote, benefits patients, either by “directly defraying the high cost of prescriptions” or by “supporting critical patient care services” [62]. The draft would let a hospital keep upfront discounts only if it passed the full value of every discount to patients. It would also require a federally set sliding fee scale, with a cap of $0 for patients below the poverty line [62, 64]. The AHA’s reply begins, “While the AHA supports efforts to help patients afford their medications,” and goes on that this approach “overlooks the purpose and structure of the 340B program” [62]. The SUSTAIN bill, which the AHA supports, would instead apply each hospital’s financial-assistance policy at its child sites [65]. The hospitals also ask Congress to bar “discriminatory practices by health plans and pharmacy benefit managers” [63]. That is a fight over who keeps the spread. If a plan pays less for a 340B prescription, part of the spread moves from the hospital to the payer. Which outcome is preferable depends on what the retained money finances and whether the purchaser’s savings reach patients or plan sponsors. Drug makers ask for fewer discounted units and more information. Manufacturers pay the discount, and the statute lets them audit a covered entity’s use of it [3]. Their preferred fixes, rebates and a narrower patient definition, would reduce the number of units sold at the ceiling price. The money from those units would stay with the manufacturer. It would lower what a patient pays only if a second rule said so. The same holds for any proposal that narrows the program, including the candidate below. Since 2020 drug makers have limited 340B sales through contract pharmacies, and two federal appeals courts have held that the statute allows it. “Nowhere does Section 340B mention contract pharmacies,” one wrote [28]. A valid concern about duplicate discounts should not be dismissed because the manufacturer benefits. But manufacturer revenue restored by a restriction is not automatically a saving for a patient or insurer. Their current vehicle is a rebate. Under an upfront discount, the entity needs less cash when it acquires the drug. Under a rebate model, it first pays more and later submits information to recover the difference. HRSA’s revised pilot is scheduled to start on January 1, 2027, for the drugs chosen for Medicare’s price negotiation [66]. A covered entity buys at the list price and claims a rebate equal to that price minus the ceiling price. The manufacturer must pay or deny with documentation within 10 days of a complete submission, and may not deny over its own concern about diversion or duplicate discounts [66]. Paying $1,000 now and receiving $400 later requires $400 more working capital than paying $600 upfront. A first version of the pilot was enjoined in December 2025 and vacated in February 2026 [67]. The AHA, the Maine Hospital Association and four hospitals challenged the revised version on October 9, 2026 [68]. The court heard argument on November 18, 2026, and no ruling had been reported by that afternoon [69, 70, 71]. HRSA audits hospitals and clinics, and it audits drug makers. It posted results for 176 audits of hospitals and clinics for fiscal 2025 [72]. In HRSA’s published table, 34 list a finding of diversion, meaning a 340B drug went to someone who was not the entity’s patient, and 78 list no adverse finding [72]. A diversion finding requires repayment to the manufacturer, and it is not a finding of fraud. HRSA chooses most of the entities it audits at random from groups it judges riskier, by purchase volume and the use of contract pharmacies among other things. It targets the rest on allegations [21, 73]. GAO reported in 2018 that the split was about 90% and 10% [21]. So these counts describe a sample chosen for risk, not a national rate. The audited hospitals are not named here, because the 34 range from large systems to rural clinics, and picking a few would accuse them of more than the audits found. HRSA also posted results for five manufacturer audits in each of fiscal 2024, 2025 and 2026 [74]. Several found that the company had “charged more than the 340B ceiling price” and ordered it to repay covered entities. The companies named are mostly generic makers [74]. One court case is unresolved. In March 2026 a federal appeals court revived a hospital system’s False Claims Act suit against AbbVie, AstraZeneca, Novartis, Sanofi and affiliates, holding only that its claims were adequately pleaded [75]. The suit alleges ceiling-price overcharges. No court has found the companies liable. The Community Oncology Alliance, COA, speaks for independent oncology practices. COA asks Congress to “Base hospital eligibility on meaningful, auditable outpatient safety-net performance” [76]. COA says it “is not asking Congress to extend 340B eligibility to independent community oncology practices” [76]. It is a member of AIR340B, a coalition that also lists PhRMA and several drug makers, and its members compete with hospitals for the same infusion patients [77]. COA’s proposal is close to mine, and COA has an interest in it. I have one too. I work in hospital oncology, in a system that takes part in 340B, and the change argued here could apply to its hospitals. HRSA can count and audit, but only Congress can change the gate. GAO examined 258 nongovernmental hospitals that HRSA reviewed in 2017 and 2018. Of them, 18 sent documents that “did not appear to be contracts” [78]. Another 13 sent contracts that did not appear to require serving the low-income population, and all were allowed to stay in the program [78]. GAO told HRSA in 2020 that it does not require hospitals to repay duplicate discounts in Medicaid managed care. The health department disagreed with two of GAO’s three recommendations, and as of March 2026 all three remained open [79]. Courts have limited HRSA from both sides. One struck down its narrow patient definition, two held that drug makers may restrict contract-pharmacy sales, and one enjoined its first rebate pilot [28, 67]. No bill before Congress changes the count of inpatient days by which a hospital qualifies. The SUSTAIN 340B Act, S. 5244, was introduced in August 2026 by Jerry Moran, Republican of Kansas [65]. Its cosponsors are Senators Tammy Baldwin of Wisconsin, Shelley Moore Capito of West Virginia, Tim Kaine of Virginia, John Boozman of Arkansas and John Hickenlooper of Colorado. Three are Democrats and two are Republicans. The bill protects upfront discounts and contract pharmacies and would end the rebate pilot. It adds a three-year wait for newly acquired sites, applies the hospital’s financial-assistance policy at child sites and requires reporting by payer [65]. It sets no test of whom a child site serves. Senator Cassidy released a discussion draft in June 2026 [80]. It lets manufacturers choose among upfront discounts, discounts after claims data, and rebates. It adds need tests for child sites and caps what patients pay, but it does not change how a hospital qualifies [64, 80]. The SECURE 340B Act, H.R. 9599, comes closest. For private nonprofit hospitals it would tighten the government-contract test, the gap GAO found, and require yearly verification of nonprofit status [81]. It is also the one bill that puts a patient test into statutory language. A child site would qualify if it lies in a socially vulnerable ZIP-code area. Failing that, it would qualify if “not less than 40 percent of patients served” are on Medicaid, uninsured or at or below 200% of the poverty line [81]. The 11.75% line stays where it is. Senator Cassidy’s draft and the SECURE bill would both bar administrators’ fees tied to the price of the drug. The draft requires flat amounts, and the bill bars fees set as a share of revenue or volume [64, 81]. All three are proposals, and a discussion draft is not enacted law. Medicare’s 2027 payment rule is final, and it pays for 340B drugs at the average sales price minus 28% [82]. It changes what Medicare pays and leaves eligibility alone. ## Measure the clinics first The first step does not depend on anyone’s preferred reform. Congress should require each 340B hospital to report whom its outpatient sites and pharmacy channels serve and what the discount earns there, on one set of definitions. Hospitals would report by site and payer for several years, which the SECURE and SUSTAIN bills already require in part, before any eligibility consequence [65, 81]. New sites would get a grace period from opening. The premise of this essay has not been tested patient by patient, and the reporting is the test. The data would be the hospital’s own audited claims, not state files. Hospitals should already be able to sort 340B claims by payer, since they bill insurers and must avoid duplicate discounts. The inspector general found in 2021 that most states’ payment data for Medicaid managed care were incomplete or inaccurate [83]. What Congress should then do with the numbers depends on what it wants the discount to pay for, and the debate runs several aims together. One is patient need, meaning how many poor patients a hospital and its sites treat. Another is uncompensated burden, the care a hospital gives and is not paid for. A third is the value of the discounted drugs that poor patients receive. These come apart. A psychiatric unit or a labor ward that serves the poor carries need and burden and uses few costly drugs. An infusion center uses a great deal of discount whoever sits in its chairs. The SECURE bill counts patients, as a fallback. A child site whose ZIP code does not qualify can still pass if 40% of its patients are on Medicaid, uninsured or poor [81]. A count like that is easy to explain and to audit. Its weakness is that a site can pass on visits while most of its drug volume is commercial. Medicare has gone the other way for its own money, toward uncompensated care. MedPAC’s index counts a hospital’s uncompensated care as a share of revenue, beside its low-income Medicare patients, inpatient and outpatient [33]. It was built to distribute Medicare payments. It is a precedent for measuring more than inpatient days, and it does not validate any 340B formula, including the one below. The entry gate’s own arithmetic could be repaired in the same direction, by counting uninsured patients in the numerator and outpatients beside inpatients. The sources would be the audited cost-report worksheet that Medicare already uses for uncompensated care and the hospital’s own claims. ## One candidate formula and its risks The candidate below weighs the drugs themselves. This is the design I would put forward. Every part of it is reasoning from the evidence above, and none of it has been tested. The test would count patients and prescriptions, never margin dollars. For each child site and each contract-pharmacy channel, count the 340B administrations and fills that go to low-income patients. Those are patients on Medicaid, uninsured, eligible for both Medicare and Medicaid, or receiving the hospital’s financial assistance. Weight each by the discount it uses, the gap between the list price and the ceiling price. The weighting matters, because a test weighted by margin would fail every hospital by construction, since Medicaid fee-for-service pays near the ceiling price. Weighting by the discount counts low-income care in the unit of the subsidy the manufacturer gives, and it counts a Medicaid patient’s infusion the same as a commercial patient’s. It also uses statuses a hospital already records, where a test of income would need data hospitals do not collect from insured patients. The reach would scale at the level of the network, with no site tested alone. A network’s 340B volume at its outpatient sites and pharmacies could run up to a set multiple of its low-income outpatient volume. Above the multiple, eligibility would taper and never stop at a cliff. This keeps the cross-subsidy, since commercial prescriptions still produce a spread, but only in proportion to the low-income drug volume behind it. An example with invented numbers shows the taper. Call a network’s low-income volume, weighted by discount, one unit, and suppose the multiple is 3. A network whose 340B volume is 2.5 units keeps all of it. Suppose that above 3 units each further unit keeps a falling share of the discount, from all of it at 3 to none at 6. A network at 5 units keeps the first 3 in full. On the next 2 it keeps two-thirds on average, the midpoint of 100% and 33%, or 1.33 units. It keeps 4.33 of its 5 units, about 87%. The multiple is a policy choice, as arbitrary as 11.75 unless evidence fixes it. > **Figure 6.** A tapered rule, one candidate among several, with an example multiple of 3. A network’s 340B volume keeps the full discount up to three times its low-income outpatient volume, then a falling share. A network at five times keeps about 87% of the discount, and none faces a cliff. Example numbers drawn for this essay; the multiple and the slope are policy choices, and how the taper would be settled is not defined. > > What the figure shows: An illustration of a candidate rule, not data. The level axis is a network’s 340B volume as a multiple of its low-income outpatient volume, and the upright axis is the volume that keeps the discount. With an example multiple of 3, the line rises one for one up to 3, so a network at 2.5 keeps all of it. Above 3 each further unit keeps a falling share, from all at 3 to none at 6, so the line bends smoothly and is level from 6 at 4.5. A network at 5 keeps 4.33, about 87 percent. A dotted line shows today’s rule, under which all of the volume keeps the discount. A gate at the clinics would wait on the reporting. If it shows that child sites serve people as needy as the parent’s, the gate is unnecessary, and I would drop it. The formula has its own perverse incentives. None has been seen, because it has never been tried. It favors expensive, deeply discounted drugs. A hospital that treats poor patients with costly infusions would earn more reach than one that gives them psychiatric, emergency, obstetric or primary care with little drug spending. That second hospital may be doing more for the poor. Drug selection and price changes would also move the allowance. A list price rises, or a pharmacy committee switches products, and a network’s reach changes while its patients are exactly as poor as before. Counting only 340B administrations and fills would undercount low-income patients whose drugs are kept out of 340B. Medicaid fee-for-service prescriptions at contract pharmacies are the plain case, since HRSA bars them by default [30]. A hospital could treat many such patients and get no credit for them. A boundary drawn around a whole network could reward acquisitions. A system could buy sites with enough qualifying volume to carry more commercial expansion elsewhere, and a rule meant to limit growth would reward buying more sites. The settlement is undefined, and it would have to be defined before anyone could vote on the rule. “Keeps the discount” could mean that the hospital keeps full discounts on chosen prescriptions. It could mean a smaller discount on every prescription, or a repayment of part of the savings after the year closes, or the loss of eligibility at chosen sites. Each lands differently on patients, on a hospital’s cash, on manufacturers and on the administrators who would run it. Site loss also brings back the cliff, one site at a time. Until the unit, the boundary of a network, the treatment of excluded prescriptions and the settlement are defined, Figure 6 illustrates an allocation rule and is not a bill. Whatever Congress chooses, rural hospitals would keep a path of their own. Critical access hospitals, sole community hospitals and rural referral centers would keep today’s eligibility, with reporting only and no new ratio. They already face no percentage test or a lower one, small volumes make ratios noisy, and a rural infusion chair serves whoever lives nearby. Joining 340B was associated with rural hospitals being 8.3 percentage points more likely to add oncology services [84]. But critical access hospitals in the most vulnerable areas are the least likely to take part at all, so a rural path does little for them [85]. Any change to the gate would also be paired with a patient rule. A low-income patient’s cost-sharing for a 340B drug would be capped at every 340B site. A gate alone moves money to manufacturers. Senator Cassidy’s $0 cap below the poverty line and SUSTAIN’s extension of financial assistance are two proposed forms [64, 65]. None of these options decides how the money is spent, which is a separate rule on the use of funds. Each adds measurement cost, and part of that cost would be paid to the administrators described above. Even audited spending allocations would leave a second problem, because money is fungible. If an institution assigns $5 million of 340B contribution to a clinic it already operated, it can redirect $5 million previously budgeted to that clinic. This is a hypothetical accounting example. ## The strongest case against me The strongest argument is that the cross-subsidy is the design, and a gate at the clinics attacks the money where it is made. Commercial patients produce the spread. It pays for a parent hospital that serves the poor and often loses money. Richmond Community is the example, in its owner’s words, and the AHA says nearly 44% of 340B hospitals run negative margins [2, 5]. A rule that asked every money-making site to serve mostly poor patients would be bound to shrink the spread, and it could close the hospitals the program keeps open. Much of that is true. In places the cross-subsidy is the reason a hospital is open. But the argument sets no limit. Under current law the reach has none, and no one can see whether the benefit keeps pace with it. For a hospital whose network really does pay for care of the poor, measuring costs only the reporting, and it lets the hospital prove its case. A tapered rule, if Congress chose one, would still let a network earn on insured patients. It would tie how far that goes to the low-income outpatients the network treats. Evidence favorable to the safety-net rationale deserves equal specificity. Newly participating public hospitals increased their offering of services that lose money, a change not seen broadly among nonprofit hospitals, whose one significant increase was in oncology [86]. Health centers that registered more 340B locations saw about 0.4% more uninsured and low-income patients for each added location, in a descriptive study [87]. All 29 covered entities GAO interviewed in 2011 said the program helped them maintain services and lower drug costs for patients, in a sample GAO called not generalizable [88]. Near the cutoff, eligibility was associated with commercial unit prices 25% lower, concentrated in competitive markets, in claims from 2012 to 2014 [89]. The findings on charity care divide. One study found a 28.9% increase in charity care after hospitals joined, and its authors concluded that the increases “appear to be fully offset by reductions in other community benefit programs” [90]. Another found no evidence that uncompensated care rose after entry [91]. A third associated new participation with about $4,075 more spending per episode of five cancer biologics among commercially insured patients. Its new entrants were mostly small and rural hospitals, and earlier participants were less likely to show the increase [92]. A scoping review of 44 studies found the same split. Its highest-quality evidence indicates that nonprofit disproportionate share hospitals “may be using the 340B program in margin-motivated ways” [93]. That finding “is not consistent across other hospital types and public health clinics” [93]. So the critical evidence is strongest for nonprofit disproportionate share hospitals, and a new gate would bind them first. It is possible for all these findings to be true. A public hospital keeps more money-losing services, such as psychiatric care, a health center expands HIV testing, a hospital system acquires a profitable infusion practice, and a pharmacy intermediary earns more fees. A new ratio costs money to compute, creates a line of its own, and invites behavior no one has studied. Washington State’s hospitals warned that a site test “incentivizes consolidation back to the main hospital campus” [94]. The AHA allows that payer-mix measures “may provide useful context” and objects to them as “rigid eligibility conditions” [95]. The current gate is volatile too, by the hospitals’ own account. The AHA warned in 2021 that pandemic changes in payer mix threatened some hospitals’ eligibility [96]. A finance association reports that roughly 425 hospitals could lose it on their next cost report [97]. The rural hospitals’ association calls 340B savings “an essential source of financial stability” and counts 417 rural hospitals as vulnerable to closure [98]. A gate alone also helps no patient. Fewer qualifying units leave more money with manufacturers. Infusion volume could return to independent practices, including those in networks that McKesson serves or that Cencora has controlled since February 2, 2026 [54, 99]. Those are the parties that gain from a narrower gate, alongside COA’s members. The timing is bad, too. In 2027 the same hospitals meet Medicare’s lower payment for 340B drugs and a faster repayment of earlier years, which the [companion essay](https://barani.org/writing/budget-neutral-340b/) works through [82]. A narrower gate would land on top of both. ## Check Two tests are possible here, and passing one proves nothing about the other. The first is whether the outpatient expansion serves a different population. Take each 340B hospital’s own claims and compute the low-income share of patients at every child site and contract-pharmacy channel. Compare it with the low-income share of the parent’s own inpatients and of the outpatients on its main campus, counted the same way. If the child sites’ share matches the parent’s, the mismatch described here belongs to Richmond and a few places like it, and it has been overstated. The second is whether the added money sustains added affordable care. A comparison of patients cannot settle that. A child site full of insured patients could be paying for a great deal of charity care, and one whose patients are as poor as the parent’s could be paying for none. This test needs net 340B revenue by site, set against the free and discounted care and the money-losing services that would shrink without it. Because money is fungible, a hospital’s own ledger will not do. The comparison has to be with similar hospitals that lack the revenue, or with the same hospital before it arrived. If care for the poor grows in step with outpatient 340B revenue, the cross-subsidy is working, and it matters much less that the inpatient gate cannot see outpatients. The data for both tests are in each hospital’s own claims and its records of financial assistance, not yet in any public file. Minnesota’s 340B reports summarize each parent entity, child sites included, rather than each site [34]. Its all-payer claims database holds claims sent by insurers and administrators, so it holds no data on care for the uninsured and has only what self-insured employer plans choose to send [100]. A payer is also a rough stand-in for a patient’s income. Some of the evidence points the other way. One analysis, sponsored by 340B Health and reported secondhand, found that 22.8% of Medicare patients receiving cancer drugs at 340B disproportionate share hospitals in 2014 were dually eligible for Medicaid. The figure in physician offices was 14.0% [101]. It comes from an interested party, and its method is not described. Still, it is patient-level and it favors hospitals. Public hospitals added money-losing services, prices were lower near the cutoff where markets were competitive, and the central case is a single hospital the committee’s majority staff examined after press reports. The public record documents violations found by audit, with repayment ordered, by hospitals and by drug makers. It also documents arrangements that are lawful and that no rule requires anyone to explain. The arithmetic can be checked. The gate counts inpatient days only, the day of an uninsured patient who is not eligible for Medicaid lowers the score, and the spread is made in outpatient settings. How far a parent’s patients differ from its child sites’ patients is less certain, because no study here measures it hospital by hospital. I am least sure what hospitals near a new line would cut. Two results would change my mind about the inpatient gate. One is patient-level data showing that child sites serve people as needy as the parent’s. The other is evidence that hospitals near a site test cut care for the poor instead of shedding affluent sites. The evidence most likely to change my assessment is audited financial information paired with credible evidence that the additional resources preserve or expand affordable care. My confidence is high that consistent definitions would improve interpretation; it is lower that disclosure alone would change conduct. Some things are missing. What any one hospital would gain or lose needs that hospital’s claims. The formula is not defined well enough to legislate. Health centers and Ryan White clinics, whose rules differ, get little space. The central case rests on one party’s staff report, and the bills and the lawsuit may move after November 14, 2026. The program’s original promise, to help scarce resources reach further, remains defensible. Its current financial architecture makes that promise difficult to verify because eligibility, revenue generation, and public benefit are measured at different points by different organizations. The question every participant should answer is therefore the same. After the drug is purchased, the patient is treated, the claim is paid, and the fees are collected, what additional care did this arrangement make possible, and for whom? ## Keep The law counts a hospital’s inpatients, but the 340B money is made on its outpatients; Congress should measure that before it moves the gate. ## Disclosure The opinions here are my own and reflect my point of view, not that of the institutions I work for, and none of them has endorsed this essay. The facts come from the public sources cited, and any correction is dated under the title. ## Sources 1. 340B Report, September 29, 2022. Fallout from the New York Times’s 340B investigative piece: a trade publication’s summary of the Times (September 24, 2022) and the Richmond Times-Dispatch, with the neighborhood, the sites’ location, the $18 million in free care and Bon Secours’s statement of compliance. [340breport.com](https://340breport.com/fallout-from-new-york-timess-340b-investigative-piece-continues-as-local-media-policymakers-and-community-weighs-in/) 2. U.S. Senate Committee on Health, Education, Labor, and Pensions, majority staff report, April 2025. *Congress Must Act to Bring Needed Reforms to the 340B Drug Pricing Program*: Richmond Community Hospital and Cleveland Clinic figures and statements, Bon Secours’s letters in the appendix, CVS and Walgreens fees and responses, and the five largest contract-pharmacy companies. One party’s staff report, not an adjudicated finding. [PDF](https://www.help.senate.gov/imo/media/doc/final_340b_majority_staff_reportpdf.pdf) 3. 42 U.S.C. § 256b: covered entities, the ceiling price, the 11.75 percent test, the other hospital routes, diversion, duplicate discounts and audits. [law.cornell.edu](https://www.law.cornell.edu/uscode/text/42/256b) 4. WTKR News 3, September 26, 2022. Virginia senator calls NYT investigation troubling: “Bon Secours needs to have an answer.” [wtkr.com](https://www.wtkr.com/news/virginia-senator-calls-nyt-investigation-troubling-bon-secours-needs-to-have-an-answer) 5. American Hospital Association, October 23, 2025. Statement to the Senate HELP Committee on the 340B program: “nearly 44% of 340B hospitals operated on a negative margin” in fiscal 2023 cost reports. An interested party’s figure. [aha.org](https://www.aha.org/testimony/2025-10-23-aha-statement-senate-help-committee-340b-program-examining-its-growth-and-impact-patients) 6. H.R. Rep. No. 102-384, pt. 2, at 9–12 (1992), as reproduced in a continuing-education packet: the best-price problem and the committee’s stated intent. [PDF](https://lgbtqbar.org/annual/wp-content/uploads/sites/6/sites/8/2024/07/340B_and_PrEP_CLE_Materials_D6.pdf) 7. 42 U.S.C. § 1396r-8(k)(1): the definition of average manufacturer price. [govinfo.gov](https://www.govinfo.gov/content/pkg/USCODE-2023-title42/html/USCODE-2023-title42-chap7-subchapXIX-sec1396r-8.htm) 8. HRSA, Office of Pharmacy Affairs Information System. Pricing formulas: “340B Ceiling Price = (AMP - URA)”; and Medicaid.gov, unit rebate amount calculation. [hrsa.gov](https://340bpricingsubmissions.hrsa.gov/Help/Manufacturer/Pricing%20Formulas/Pricing%20Formulas.htm), [medicaid.gov](https://www.medicaid.gov/medicaid/prescription-drugs/medicaid-drug-rebate-program/unit-rebate-amount-calculation) 9. Nikpay, Reinke & Quinones, *Milbank Quarterly*, 2026. A legislative history of 340B from 175 internal documents and 19 interviews; funded by the Commonwealth Fund. [doi.org](https://doi.org/10.1111/1468-0009.70090) 10. Congressional Budget Office, September 9, 2025. *Growth in the 340B Drug Pricing Program*: purchases of $6.6 billion (2010) and $43.9 billion (2021), the 87% outpatient share, integration as the largest factor, and no statutory rule on use of revenue. [cbo.gov](https://www.cbo.gov/publication/60661), [PDF](https://www.cbo.gov/system/files/2025-09/60661-340B-program.pdf) 11. HRSA HIV/AIDS Bureau, March 21, 2016. Frequently asked questions on Policy Clarification Notices 15-03 and 15-04: 340B revenue as program income and its permitted uses. [PDF](https://www.hab.hrsa.gov/sites/default/files/ryanwhite/grants/faq-15-03-15-04.pdf) 12. HRSA. 2025 340B covered entity purchases, accurate as of June 30, 2026: $100.0 billion in total and $79.2 billion by disproportionate share hospitals. [hrsa.gov](https://www.hrsa.gov/opa/updates/2025-340b-covered-entity-purchases) 13. HRSA. Disproportionate share hospitals: eligibility on the most recently filed cost report; for-profit hospitals are not eligible. [hrsa.gov](https://www.hrsa.gov/opa/eligibility-and-registration/hospitals/disproportionate-share-hospitals) 14. Centers for Medicare & Medicaid Services. Disproportionate share hospital (DSH): the patient-percentage formula and the 25%/75% split of DSH money since the Affordable Care Act. [cms.gov](https://www.cms.gov/medicare/medicare-fee-for-service-Payment/AcuteInpatientPPS/dsh) 15. 42 C.F.R. § 412.106: which days count, the adjustment formula for urban hospitals of 100 or more beds, and the 12 percent cap. [law.cornell.edu](https://www.law.cornell.edu/cfr/text/42/412.106) 16. Medicare Payment Advisory Commission, May 2015. *Report to the Congress: Overview of the 340B Drug Pricing Program*: outpatient sites do not affect the percentage; DSH payments as a proxy for uncompensated care. [PDF](https://www.medpac.gov/wp-content/uploads/./import_data/scrape_files/docs/default-source/reports/may-2015-report-to-the-congress-overview-of-the-340b-drug-pricing-program.pdf) 17. U.S. House Committee on Energy and Commerce, Subcommittee on Oversight and Investigations, October 11, 2017. *Examining How Covered Entities Utilize the 340B Drug Pricing Program*, Serial No. 115-63: testimony of Northside Hospital’s director of finance. [govinfo.gov](https://www.govinfo.gov/content/pkg/CHRG-115hhrg27577/html/CHRG-115hhrg27577.htm) 18. HRSA, October 27, 2023. Notice on registration of off-site outpatient facilities, 88 Fed. Reg.: the cost-report test for child sites. [PDF](https://www.govinfo.gov/content/pkg/FR-2023-10-27/pdf/2023-23702.pdf) 19. HRSA. Frequently asked question: can non-Medicaid patients receive 340B drugs? [hrsa.gov](https://www.hrsa.gov/es/node/844) 20. HRSA, October 24, 1996. Notice regarding the 340B patient and entity eligibility definition, 61 Fed. Reg. 55156. [govinfo.gov](https://www.govinfo.gov/content/pkg/FR-1996-10-24/html/96-27344.htm) 21. U.S. Government Accountability Office, June 2018. GAO-18-480: contract pharmacies from about 1,300 (2010) to nearly 20,000 (2017); how HRSA selects entities for audit; discounts to uninsured patients at 30 of 55 entities reviewed. [gao.gov](https://www.gao.gov/products/gao-18-480) 22. *AbbVie v. Wrigley*, No. 1:25-cv-00182 (D.N.D. April 27, 2026), order at 7–9: the replenishment model, described from a manufacturer’s declaration. [govinfo.gov](https://www.govinfo.gov/content/pkg/USCOURTS-ndd-1_25-cv-00182/pdf/USCOURTS-ndd-1_25-cv-00182-2.pdf) 23. Healthcare Dive, September 2025. A report of CBO’s figures on off-site clinics, 2013 and 2021. [healthcaredive.com](https://www.healthcaredive.com/news/cbo-340B-report-hospitals-spending/759677/) 24. Zeng, Sarraille & Martin, *Health Affairs Scholar* 3:qxaf104, 2025. What is driving 340B growth: utilization or price? Partly supported by the National Pharmaceutical Council; editorial control retained by IQVIA; data not public. [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC12188212/) 25. National Pharmaceutical Council. Members: “supported by leading biopharmaceutical companies.” [npcnow.org](https://www.npcnow.org/about/members) 26. Advis, February 15, 2021. Establish medication management clinic to expand 340B benefit. A consultant’s marketing. [advis.com](https://advis.com/health-smart/healthcare-revenue-tips2021/) 27. CPS. 340B referral capture. A vendor’s marketing, undated, read in November 2026. [cps.com](https://discover.cps.com/340B-Referral-Capture) 28. Congressional Research Service, September 10, 2025, R48696 (archival copy): *Sanofi-Aventis v. HHS* (3d Cir.), *Novartis v. Johnson* (D.C. Cir.) and *Genesis Healthcare v. Becerra*. [everycrsreport.com](https://www.everycrsreport.com/files/2025-09-10_R48696_b70a409b099de5e485938161afd1a9b299edf90e.html) 29. CMS, February 11, 2016. State Health Official letter #16-001: Medicaid payment at actual acquisition cost plus a dispensing fee; 340B drugs not above the ceiling price. [PDF](https://www.medicaid.gov/federal-policy-guidance/downloads/smd16001.pdf) 30. HRSA. Frequently asked question on Medicaid fee-for-service patients at contract pharmacies. [hrsa.gov](https://www.hrsa.gov/es/node/847) 31. HRSA, March 5, 2010. Notice regarding 340B contract pharmacy services, 75 Fed. Reg. 10272. [govinfo.gov](https://www.govinfo.gov/content/pkg/FR-2010-03-05/html/2010-4755.htm) 32. HHS Office of Inspector General, February 4, 2014. OEI-05-13-00431, Contract pharmacy arrangements in the 340B program: a purposive sample of 30 covered entities. [oig.hhs.gov](https://oig.hhs.gov/reports/all/2014/contract-pharmacy-arrangements-in-the-340b-program/) 33. Medicare Payment Advisory Commission, March 2023. *Report to the Congress*, chapter 3: the DSH formula and the three parts of the Medicare Safety-Net Index. [PDF](https://www.medpac.gov/wp-content/uploads/2023/03/Ch3_Mar23_MedPAC_Report_To_Congress_SEC.pdf) 34. Minnesota Department of Health, February 27, 2026. *340B Covered Entity Report, 2025* (2024 data): payments, acquisition and operating costs, net revenue, payer estimates and limitations. Self-reported and unaudited. [PDF](https://www2cdn.web.health.state.mn.us/data/340b/docs/2025report.pdf) 35. Maine Revised Statutes, Title 22, § 1728, subsection 2-A (P.L. 2025, c. 388): annual 340B hospital reports to the Maine Health Data Organization beginning July 1, 2026, by payor type. [legislature.maine.gov](https://legislature.maine.gov/statutes/22/title22sec1728.html) 36. Washington State Legislature, Engrossed Second Substitute Senate Bill 5981, Chapter 227, Laws of 2026, effective June 11, 2026, section 5: annual 340B reports to the Health Care Authority, by payer type where available. [PDF](https://lawfilesext.leg.wa.gov/biennium/2025-26/Pdf/Bills/Senate%20Passed%20Legislature/5981-S2.PL.pdf) 37. Robinson, Kosorukov & Whaley, *JAMA* 335:1243–1249, 2026. Hospital adoption and pricing for oncology biosimilars: Blue Cross Blue Shield claims, 2020–2024; acquisition prices estimated; supported by the NIHCM Foundation. [jamanetwork.com](https://jamanetwork.com/journals/jama/fullarticle/2846346) 38. NIHCM Foundation. Board of directors. [nihcm.org](https://nihcm.org/about-us/board-of-directors) 39. 340B Report, January 2023. MedPAC votes to recommend ditching Medicare DSH patient percentage for Medicare purposes. Trade press, the source of the director’s quotation. [340breport.com](https://340breport.com/medpac-votes-to-recommend-ditching-medicare-dsh-patient-percentage-for-medicare-purposes/) 40. Desai & McWilliams, *New England Journal of Medicine* 378:539–548, 2018. Consequences of the 340B Drug Pricing Program: regression discontinuity at the 11.75 percent threshold, Medicare claims. [doi.org](https://doi.org/10.1056/NEJMsa1706475) 41. American Hospital Association, February 5, 2018. Major methodological flaws negate findings of recent 340B study. An interested party’s reply. [aha.org](https://www.aha.org/news/blog/2018-02-05-major-methodological-flaws-negate-findings-recent-340b-study) 42. Mulligan, Romley & Myerson, *BMC Research Notes* 14, 2021. Hospitals just above and below the 340B threshold, 2014–2016, with investor-owned hospitals as a placebo. [doi.org](https://doi.org/10.1186/s13104-021-05642-4) 43. Thome, Odouard & Buntin, *JAMA Health Forum* 7(10):e263977, 2026. Hospital 340B participation, site-dependent payment, and vertical integration: difference in differences, 122 new 340B hospitals against 779 others, traditional Medicare claims 2016–2019. [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC13723016) 44. Conti & Bach, *Health Affairs* 33(10), 2014. The 340B drug discount program: hospitals generate profits by expanding to reach more affluent communities. 960 hospitals and 3,964 clinics. [doi.org](https://doi.org/10.1377/hlthaff.2014.0540) 45. Liu, Russo, Kesselheim, Knox, Sarpatwari & Feldman, *Health Services Research*, 2025. Serial cross-section of newly registered hospitals, child sites and contract pharmacies, 2010–2022, by a social deprivation index. [doi.org](https://doi.org/10.1111/1475-6773.14446) 46. Masia, Filson, Martin & Neumann, *Health Affairs Scholar*, 2025. Child-site locations in 2022; funded by Johnson & Johnson. [doi.org](https://doi.org/10.1093/haschl/qxaf121) 47. 340B Informed. Hospitals rebut Health Affairs study’s 340B conclusions: the reply of the hospitals’ association, now 340B Health. A hospital-aligned outlet. [340bemployed.org](https://340bemployed.org/hospitals-rebut-health-affairs-studys-340b-conclusions) 48. PhRMA, February 2025. *Follow the Dollar*: wholesalers, chargebacks, and who benefits from a 340B transaction. An interested party’s report. [PDF](https://cdn.aglty.io/phrma/global/resources/import/pdfs/Report%20-%20Follow%20the%20Dollar%20-%20February%202025.pdf) 49. Gabler, *New York Times*, February 12, 2026, as republished by GV Wire. Senate questions health care firm for profiting off program meant for poor. [gvwire.com](https://gvwire.com/2026/02/12/senate-questions-health-care-firm-for-profiting-off-program-meant-for-poor) 50. Senator Bill Cassidy, chairman, Senate HELP Committee, February 1, 2026. Letter to Apexus. A senator’s questions, not findings. [PDF](https://www.help.senate.gov/imo/media/doc/26-02-01_chairman_cassidy_letter_to_apexus_finalpdf.pdf) 51. Becker’s Hospital Review. 340B contractor boosts profits while driving up patient costs: report. The source of Apexus’s statement. [beckershospitalreview.com](https://beckershospitalreview.com/pharmacy/340b-contractor-boosts-profits-while-driving-up-patient-costs-report.html) 52. Cencora, Form 10-K for the fiscal year ended September 30, 2025. [PDF](https://s203.q4cdn.com/785768684/files/doc_downloads/2026/01/cor-20250930.pdf) 53. McKesson. Protecting hospitals and patients with 340B: Macro Helix’s clients. The company’s own account. [mckesson.com](https://www.mckesson.com/stories-insights/protecting-hospitals-and-patients-with-340b/) 54. McKesson, June 2025. McKesson Corporation completes acquisition of Core Ventures. [mckesson.com](https://www.mckesson.com/about-us/newsroom/press-releases/2025/mckesson-corporation-completes-acquisition-of-core-ventures/) 55. CVS Health, Form 10-K for the fiscal year ended December 31, 2025. [PDF](https://s206.q4cdn.com/752775519/files/doc_financials/2025/q4/2025-Form-10-K.pdf) 56. Pharmacy services agreement between AIDS Service Association of Pinellas (EPIC) and Community Specialty Pharmacy, effective April 1, 2021, Exhibit 10.22 to a 2024 SEC filing. [sec.gov](https://www.sec.gov/Archives/edgar/data/2030763/000149315224049098/ex10-22.htm) 57. Federal Trade Commission, July 9, 2024. Interim staff report on pharmacy benefit managers. [ftc.gov](https://www.ftc.gov/news-events/news/press-releases/2024/07/ftc-releases-interim-staff-report-prescription-drug-middlemen) 58. Lincoln International. Verity Solutions has been sold to Express Scripts, a division of Cigna. A deal adviser’s notice. [lincolninternational.com](https://www.lincolninternational.com/transactions/verity-solutions-has-been-sold-to-express-scripts-a-division-of-cigna/) 59. PhRMA, lobbying report (LD-2), third quarter of 2026: $7,040,000 across all issues. [lda.gov](https://lda.gov/filings/public/filing/afab4741-599f-4aea-ae8f-597ba4797e76/print/) 60. American Hospital Association, lobbying report (LD-2), third quarter of 2026: $3,840,000 across all issues. [lda.gov](https://lda.gov/filings/public/filing/5f715aae-6fa2-41fd-b465-39e61cd8a3f2/print/) 61. 340B Health, lobbying report (LD-2), third quarter of 2026: $330,000 across all issues. [lda.gov](https://lda.gov/filings/public/filing/86116d8d-d5f8-4865-9237-a95238e60523/print/) 62. American Hospital Association, August 25, 2026. Feedback on the 340B Drug Pricing Integrity and Affordability for Patients Act discussion draft. [aha.org](https://www.aha.org/letterscomments/2026-08-25-aha-feedback-re-340b-drug-pricing-integrity-and-affordability-patients-act-340b-patients-act) 63. American Hospital Association, October 27, 2026. Comments on the SUSTAIN 340B Act, S. 5244, to its six sponsors. [aha.org](https://www.aha.org/letterscomments/2026-10-28-aha-comments-sustain-340b-act-s-5244) 64. Academy of Managed Care Pharmacy, July 14, 2026. Summary of the Cassidy discussion draft and the SECURE 340B Act: the $0 cap below the poverty line, child-site tests and flat administrator fees. A trade group’s summary. [PDF](https://www.amcp.org/sites/default/files/2026-07/legupdate_summaryofcassidy_peters340Bproposals.pdf) 65. S. 5244, SUSTAIN 340B Act, 119th Congress, introduced August 5, 2026. [govinfo.gov](https://www.govinfo.gov/content/pkg/BILLS-119s5244is/pdf/BILLS-119s5244is.pdf) 66. HRSA, August 3, 2026. 340B Rebate Model Pilot Program, revised notice, 91 Fed. Reg. 48883. [govinfo.gov](https://www.govinfo.gov/content/pkg/FR-2026-08-03/html/2026-15633.htm) 67. National Law Review, February 25, 2026. 340B rebate model pilot program effectively ends, preserving upfront discounts for now. [natlawreview.com](https://natlawreview.com/article/340b-rebate-model-pilot-program-effectively-ends-preserving-upfront-discounts-now) 68. *American Hospital Association v. Kennedy*, No. 1:26-cv-00515 (D. Me.), plaintiffs’ motion, October 2026. Allegations and legal arguments, not judicial findings. [PDF](https://www.aha.org/system/files/media/file/2026/10/AHA-v-Kennedy-340B-Temporary-Restraining-Order-20261008.pdf) 69. *American Hospital Association v. Kennedy*, No. 1:26-cv-00515 (D. Me.), notice of hearing, November 6, 2026: oral argument set for November 18, 2026. Court document reproduced by Georgetown’s Health Care Litigation Tracker. [PDF](https://litigationtracker.law.georgetown.edu/wp-content/uploads/2026/10/AHA-II_2026.11.06_SCHEDULING-ORDER.pdf) 70. Bloomberg Law, November 18, 2026. Judge stays quiet in fight over Trump’s latest drug rebate plan. A report of the hearing. [news.bloomberglaw.com](https://news.bloomberglaw.com/us-law-week/judge-stays-quiet-in-fight-over-trumps-latest-drug-rebate-plan) 71. 340B Report, November 18, 2026. Fed judge holds hearing on fate of HRSA’s revamped 340B rebate pilot. A report of the hearing. [340breport.com](https://340breport.com/fed-judge-holds-hearing-on-fate-of-hrsas-revamped-340b-rebate-pilot/) 72. HRSA. Program integrity: fiscal year 2025 audit results for covered entities. [hrsa.gov](https://hrsa.gov/opa/program-integrity/fy-25-audit-results) 73. HRSA, February 8, 2013. Clarification of HRSA audits of 340B covered entities, Release No. 2012-1.1: risk-based and targeted audits. [PDF](https://www.hrsa.gov/sites/default/files/hrsa/opa/audit-clarification-02-08-13.pdf) 74. HRSA. Program integrity: manufacturer audit results for fiscal years 2024, 2025 and 2026. [FY 2024](https://www.hrsa.gov/opa/program-integrity/fy-24-manufacturer-audit-results), [FY 2025](https://www.hrsa.gov/opa/program-integrity/fy-25-manufacturer-audit-results), [FY 2026](https://hrsa.gov/opa/program-integrity/fy-26-manufacturer-audit-results) 75. *United States ex rel. Adventist Health System/West v. AbbVie Inc.*, No. 24-2180 (9th Cir. March 17, 2026): reversal of a dismissal for failure to state a claim. [PDF](https://cdn.ca9.uscourts.gov/datastore/opinions/2026/03/17/24-2180.pdf) 76. Community Oncology Alliance, August 19, 2026. Announcement of its recommendations on Senator Cassidy’s discussion draft, distributed by EIN Presswire. An interested party’s statement. [einpresswire.com](https://www.einpresswire.com/article/935401219/coa-submits-patient-centered-recommendations-on-sen-cassidy-s-340b-reform-discussion-draft) 77. Alliance for Integrity and Reform of 340B. Members. [340breform.org](https://340breform.org/members/) 78. U.S. Government Accountability Office, December 2019, released January 2020. GAO-20-108: nongovernmental hospitals’ contracts. [gao.gov](https://www.gao.gov/products/gao-20-108) 79. U.S. Government Accountability Office, January 2020, with recommendation status through March 2026. GAO-20-212: duplicate discounts in Medicaid managed care. [gao.gov](https://www.gao.gov/products/gao-20-212) 80. Senate HELP Committee, June 2026. Section-by-section summary of the 340B Drug Pricing Integrity and Affordability for Patients Act discussion draft. [help.senate.gov](https://www.help.senate.gov/sxs-340b-drug-pricing-integrity-and-affordability-for-patients-act-discussion-draftpdf) 81. H.R. 9599, SECURE 340B Act, 119th Congress, introduced July 6, 2026. [govinfo.gov](https://www.govinfo.gov/content/pkg/BILLS-119hr9599ih/xhtml/BILLS-119hr9599ih.html) 82. CMS, November 4, 2026. Fact sheet, CY 2027 OPPS and ASC final rule: payment for 340B drugs at the average sales price minus 28% and the faster repayment. [cms.gov](https://www.cms.gov/newsroom/fact-sheets/calendar-year-2027-hospital-outpatient-prospective-payment-system-opps-ambulatory-surgical-center-0) 83. HHS Office of Inspector General, March 2021. OEI-02-19-00180: states’ Medicaid managed-care payment data. [oig.hhs.gov](https://oig.hhs.gov/oei/reports/OEI-02-19-00180.asp) 84. Owsley & Bradley, *Health Affairs* 42(6), 2023. Rural hospitals’ 340B participation and oncology services, 2011–2020. [doi.org](https://doi.org/10.1377/hlthaff.2022.01640) 85. Owsley & Karim, *Journal of Rural Health*, 2024. Critical access hospitals, community vulnerability and 340B participation. [doi.org](https://doi.org/10.1111/jrh.12833) 86. Owsley and colleagues, *JAMA Health Forum*, 2024. 340B participation and unprofitable services: difference in differences, 2,152 hospitals, 2010–2019. [doi.org](https://doi.org/10.1001/jamahealthforum.2024.0833) 87. Watts and colleagues, *JAMA Health Forum*, 2024. 340B registered locations at federally qualified health centers: descriptive retrospective cohort. [jamanetwork.com](https://jamanetwork.com/journals/jama-health-forum/fullarticle/2824253) 88. U.S. Government Accountability Office, September 2011. GAO-11-836: interviews with 29 covered entities, a judgmental sample. [gao.gov](https://www.gao.gov/assets/a323707.html) 89. Desai and colleagues, *Health Services Research*, 2026. 340B eligibility and negotiated drug prices: regression discontinuity, commercial claims 2012–2014. [doi.org](https://doi.org/10.1111/1475-6773.70085) 90. Nikpay, Buntin & Conti, *Health Services Research* 55(2), 2020. 340B participation, charity care and community benefit. [doi.org](https://doi.org/10.1111/1475-6773.13278) 91. Desai & McWilliams, *American Journal of Managed Care* 27(10):432–437, 2021. 340B and hospital provision of uncompensated care, 2003–2015. [doi.org](https://doi.org/10.37765/ajmc.2021.88761) 92. Chang and colleagues, *JAMA Health Forum*, 2023. Hospital 340B participation and spending on five oncology biologics, commercially insured patients, 2007–2019. [jamanetwork.com](https://jamanetwork.com/journals/jama-health-forum/fullarticle/2806517) 93. Levengood and colleagues, *Milbank Quarterly*, 2024. A scoping review of 44 studies of the 340B program. [doi.org](https://doi.org/10.1111/1468-0009.12691) 94. Washington State Hospital Association, August 27, 2026. Comment letter on Senator Cassidy’s discussion draft. An interested party’s statement. [PDF](https://www.wsha.org/webfoo/wp-content/uploads/WSHA-Comment-Letter_340B-Sen.-Cassidy-Response-Final.pdf) 95. American Hospital Association, September 23, 2026. Comments on the SECURE 340B Act. [aha.org](https://www.aha.org/letterscomments/2026-09-23-aha-comments-secure-340b-act) 96. American Hospital Association, September 2021. Fact sheet: 340B-related payer mix changes. [PDF](https://www.aha.org/system/files/media/file/2021/09/fact-sheet-340b-related-payer-mix-changes-9-2021.pdf) 97. Healthcare Financial Management Association, updated September 29, 2026. Continued 340B eligibility is at risk for hundreds of hospitals, citing rural-hospital association data. [hfma.org](https://www.hfma.org/payment-reimbursement-and-managed-care/medicare-cost-report/continued-340b-eligibility-is-at-risk-for-hundreds-of-hospitals/) 98. National Rural Health Association, August 28, 2026. Comment on the 340B reform discussion draft. An interested party’s statement. [PDF](https://www.ruralhealth.us/NationalRuralHealth/media/Documents/Advocacy/2026/NRHA-340B-Reform-Comment.pdf) 99. Cencora, Form 10-Q for the quarter ended March 31, 2026. Acquisition of the majority of OneOncology, completed February 2, 2026. [sec.gov](https://www.sec.gov/Archives/edgar/data/0001140859/000114085926000020/cor-20260331.htm) 100. Minnesota House Research, August 2026. Minnesota All Payer Claims Database: sources of data; self-insured ERISA plans report voluntarily; and Minnesota Department of Health, About the MN APCD: no data on care for the uninsured. [PDF](https://www.house.mn.gov/hrd/pubs/ss/ssmnapcd.pdf), [health.state.mn.us](https://www.health.state.mn.us/data/apcd/about/index.html) 101. 340B Informed, May 2017. 340B hospitals treat more low-income cancer patients than non-340B providers: a Dobson DaVanzo analysis for 340B Health, 2014 Medicare Part B data, reported secondhand. [340binformed.org](https://340binformed.org/2017/05/340b-hospitals-treat-more-low-income-cancer-patients-than-non-340b-providers) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/met-the-minutes/ Markdown: https://barani.org/writing/met-the-minutes/index.md Published: 2026-11-12 Modified: 2026-11-12 Form: Essay Topic: Mathematics teaching Series: How children learn mathematics, part 2 of 4 ======================================================================== # Children who met their math minutes gained more. What does that prove? A green dashboard, a spring report and the comparison a school needs instead. Igor Barani · 12 November 2026 · 16 min read Mathematics teaching · Essay · How children learn mathematics, part 2 of 4 · https://barani.org/writing/met-the-minutes/ A school’s math dashboard turns green when a child logs her 30 minutes for the week. In spring a report shows that the children who met the target gained more than the children who did not, so the school resolves to get every child to green. The report cannot carry the school’s resolution. It shows that children who meet the target differ from children who do not. It cannot show that requiring the same minutes of the children who missed it would raise their scores, because nobody assigned the minutes. The children, their teachers and their circumstances chose them. I think a school learns what a met target did only from a comparison in which the minutes themselves were assigned, by lot where possible. I would keep the target as a floor that shows the program reached the child, and stop counting minutes beyond it as progress. What chooses the minutes can also move the scores. A teacher who schedules the sessions, a child who comes to school every day and a home that checks the work can each add minutes. Each can also add learning of its own. So a report that sorts children by minutes also sorts them by these causes, and a rule that requires the minutes need not bring the causes with it. I am writing With Every Reason, a series of arithmetic books being prepared for printing and for adoption by school districts, with a materials kit sold separately. Print books compete with software for children’s minutes, so an essay that doubts software targets serves my interest. The standard I argue for binds my books too, because a finished workbook proves no more than a met target. No trial has yet tested my books. ## What the i-Ready brief shows In June 2021 Curriculum Associates, the company that sells i-Ready, an online math and reading program, published a research brief on time and learning [1]. It sorted 1,167,430 students, from kindergarten to eighth grade, by their use of i-Ready math in the 2017–2018 school year. It made four groups and analyzed three. It left out the fourth, children who averaged 1 to 9 minutes a week and passed fewer than 70% of their quizzes, because it held only 106 children. Children in the group the company recommends averaged 30 to 49 minutes a week and passed at least 70% of their lesson quizzes. A low group averaged 10 to 29 minutes and passed fewer than 70%. A high group averaged 50 minutes or more and passed at least 70%. None of the four groups has a place for a child who logged many minutes and failed often, or the reverse. The measure was the gain from fall to spring on the i-Ready Diagnostic, the company’s own test. Third graders in the low group gained about 25 points on its scale, and those in the recommended group about 30. The brief sizes that gap at 0.22 of the usual spread among students, roughly how far a typical child sits from the average. Researchers call that spread a standard deviation, and the sections below measure in it. The gap was about five to seven points in every grade. The brief presents this as evidence for using the program as recommended [1]. Its limitations section adds that the study “is observational in nature” and that its inferences “are correlational and not causal” [1]. A school that reads the brief as a reason to require minutes takes a step that section declines. Two features of the comparison show why the step is unsafe. The groups differ in two ways at once, since the recommended group logged more minutes and also passed more quizzes. A child who passes most of her quizzes is already learning from the lessons. So the groups were sorted partly by the thing the study set out to measure. The low group is also a narrow slice. It held about 59,000 children, one in twenty of those studied, who both logged few minutes and passed fewer than 70% of their quizzes. So the comparison sets ordinary users against children already in trouble, and the gap measures that difference along with the minutes. ## A sugar pill shows the trap A heart trial held a control that schools seldom have. The Coronary Drug Project assigned men who had survived a heart attack, at random, to a cholesterol drug called clofibrate or to a placebo, an inert pill. After five years, 20.0% of the 1,103 men on the drug had died and 20.9% of the 2,789 men on placebo [2]. Chance could explain a difference that small, so the drug had done little or nothing. Then the researchers split the drug group by how faithfully the men had taken it. Among those who took at least 80% of their prescribed pills, 15.0% died. Among the rest, 24.6% died [2]. Read alone, that gap says the drug saves lives in the men who take it. The placebo group, split the same way, showed a wider gap, 13 points against 10. Of the men who took at least 80% of their inert pills, 15.1% died, against 28.3% of the rest [2]. No one concluded that sugar pills prevent death. The men who took their pills faithfully differed from the men who did not, in their health, their habits and their circumstances. Those differences, and nothing in the pill, account for the gap. Assignment by lot had made the two arms of the trial alike. Splitting an arm afterward by what the men did sorted them again, this time by everything that makes a man take his pills. > **Figure 1.** Deaths within five years in the Coronary Drug Project, by assigned pill and by how faithfully it was taken. As assigned, the two pills barely differ, 20.0% against 20.9%. Within each arm the men who took at least 80% of their pills died far less often. The gap is at least as wide on the inert pill, 13 points against 10, so it comes from the men and not from the pills. Data from the 1980 report [2]. > > What the figure shows: Two pairs of bars for deaths within five years in the Coronary Drug Project. Among all men assigned the drug, 20.0 percent died, against 20.9 percent of all men assigned the inert placebo, so the pills compared as assigned barely differ. Within the drug arm, 15.0 percent of men who took at least 80 percent of their pills died, against 24.6 percent of those who took fewer. Within the placebo arm the figures were 15.1 and 28.3 percent. The gap between faithful and unfaithful pill-takers is at least as wide for the placebo as for the drug. The trial holds a second lesson, for anyone told that a report controlled for prior scores. The 1980 researchers adjusted for 40 of the men’s characteristics at the start of the trial, and a gap of about 10 points remained [3]. Decades later, two epidemiologists reanalyzed the placebo group with methods that also account for what happened to each man during the five years. The gap shrank to between −0.7 and 4.5 points, depending on the method, and each estimate was uncertain enough to include zero [4]. A man who grows sicker may stop his pills, and he is also more likely to die. A snapshot taken at the start cannot hold that. The i-Ready analysis adjusted for each child’s fall scores, and the brief describes no other adjustment [1]. A child’s fall score is the same kind of snapshot. What follows it, a long illness, a new teacher, a concept that finally lands, moves her minutes and her spring score together. ## The same test, run on a math program A school has no placebo software, but it has something close. A math program should do little for reading. So if children who log more math minutes also gain more in reading, the minutes probably did not make that gain. Researchers at Khan Academy ran this test on their own product. They studied 181,359 students in grades 3 to 8, in 649 schools that had adopted a math tool called MAP Accelerator, in the 2020–2021 school year [5]. Teachers were encouraged to give each child at least 30 minutes a week, and 5% of the children reached that. Math growth rose with each level of use. Then the authors looked at reading and found “a similar pattern” there. They called it “problematic as it suggests the existence of unobserved confounding in the usage groups” [5]. They subtracted the reading pattern from the math one and reported what remained. Growth was higher at 30 minutes or more than at under 15, by 0.26 standard deviations, about a quarter of that spread. The subtraction removes whatever lifts both subjects alike, such as attendance or a parent who checks homework. It cannot remove what lifts math alone, such as a strong math teacher. And if the math tool did help reading a little, the subtraction takes away too much. The same paper measured where minutes come from. Teacher, school and district together explained 57% of the differences in how many minutes a child logged [5]. So a child’s minutes are in large part a fact about her classroom. A teacher who gets every child logged in each week has probably organized other things as well, and her students’ scores carry all of it. The bias need not flatter the program. A teacher may send the children who are falling behind to the software for extra time, and a child who understands quickly may finish in fewer minutes. Either habit would make heavy users look worse than the program made them. ## Three questions hide in “it works” A school that asks whether a program works is asking one of three questions, and each needs its own evidence. The adoption question is whether taking up the program raises scores. A trial that assigns some schools the program and leaves others without it answers that. The amount question is whether more minutes raise scores further. Only a trial that assigns different amounts answers that. The usage question is whether the children who use the program more score higher. A usage report answers that one and no other. A minutes target is a claim about the amount question. Evidence on the adoption question, however strong, is silent on it. ## A lottery for the program is not a lottery for the minutes Zearn Math shows how far apart the adoption and usage questions can sit. According to the RAND report described below, Zearn recommends 90 minutes and three lessons a week. The same report describes Zearn’s own studies. They matched students who met usage goals with similar students who did not, and found the first group ahead by 0.27 to 0.40 standard deviations on three tests [6]. Across 1,942 effects from 747 randomized trials of school programs, the median was 0.10 standard deviations. By that measure, 0.27 to 0.40 would be three to four times the usual effect, if the program caused it [7]. Researchers at RAND designed a trial with Zearn and ran it by lottery on a federal grant. They assigned 64 schools in one Texas district at random, half to use Zearn and half to carry on as before, other software included, and followed about 10,600 students for two years. On the state test the Zearn schools finished 0.07 standard deviations ahead, with an interval from −0.04 to 0.19 that includes zero [6]. The authors wrote that the earlier differences were likely “upwardly biased,” because “students who use Zearn Math more may be more motivated” [6]. Yet the trial does not refute the usage studies. In its first year 18% of the Zearn students completed the recommended lessons, and in its second year 50% did [6]. Suppose, as an example, that Zearn raised those students by 0.33, the middle of the usage studies’ range, and the rest by nothing. The average over everyone offered it would be 0.06 to 0.17, which sits inside the trial’s interval. So does zero. The lottery assigned the program, and the children still chose their minutes. So the trial answers the adoption question and leaves the amount question open. ## One trial assigned the amount One large trial assigned two amounts of the same software. In Russia, researchers took one third-grade class in each of 343 schools and assigned the classes by lot to three groups for five months [8]. One group kept its ordinary homework. A second had some or all of that homework replaced by 10 software exercises a week in math and 10 in language, about 20 to 25 minutes for each subject. A third got 20 exercises a week in each, about 40 to 50 minutes. Both software groups gained in math. Against the classes with ordinary homework, the classes on 10 exercises a week scored 0.099 standard deviations higher and the classes on 20 exercises a week 0.087 [8]. The second 10 exercises bought nothing that the trial could detect. The two software groups differed by −0.012 standard deviations, and the 95% interval for that difference runs from about −0.09 to 0.06, so a small loss and a small gain both fit. Had doubling the work doubled the gain, the classes on 20 exercises would have gained about 0.20. The paper’s own test puts them 0.111 short of that. The shortfall rests on two estimates, and its standard error, which measures how uncertain it is, is 0.067. So at the usual 95% standard the trial cannot rule out that doubling the work doubled the gain [8]. > **Figure 2.** Math gains in the Russian trial for classes assigned 10 or 20 software exercises a week, measured from classes that kept ordinary homework. Both gained about a tenth of a standard deviation, and the two intervals nearly coincide. The dotted line marks twice the first gain, itself an estimate, so the distance to it is less sure than it looks. Estimates from the trial [8], with 95% intervals computed for this essay from its standard errors. > > What the figure shows: Two estimates from the Russian trial, in standard deviations of math scores, measured from classes that kept ordinary homework. Ten exercises a week gained 0.099, with a 95 percent interval from 0.023 to 0.175. Twenty exercises a week gained 0.087, with an interval from 0.011 to 0.163. A dotted line at 0.198 marks twice the first gain. The doubled workload lies below it, though twice the first gain is itself an estimate and the trial cannot rule out a proportional gain at the 95 percent standard. Both estimates sit above zero. The result has other limits. The trial assigned exercises, and the minutes are estimates, because the software’s provider could not supply complete usage records [8]. The trial’s authors thank Yandex for data and support, but the paper does not name the software’s provider, so this essay cannot say whose product was tested [8]. And these were third graders doing homework in two Russian provinces, so the result is no ruling on an American classroom. A search on November 11 and 12, 2026, found no randomized trial in American schools that assigned different weekly minutes of one software program. One small trial of a school’s tutoring varied the sessions a week together with the size of the group [9]. ## The minutes come from somewhere If a second batch of exercises, another 20 to 25 minutes a subject, bought little that a trial could see, what the minutes cost matters more. Thirty minutes a week over a 36-week school year is 18 hours, and the hours come out of something. Software that replaces a worksheet is one trade, and software that replaces a teacher’s small group is another. A trial in rural China made the trade explicit. About 4,000 boarding students in grades 4 to 6 took part, and their classes were assigned by lot to three groups. One had a weekly 40-minute session of math software, one had the same session in a paper workbook, and one had none [10]. On the math test, the software group stood 0.033 standard deviations above the group with no session and 0.059 above the workbook group. Neither difference could be told from zero [10]. The workbook group scored about the same as the software group. I make print books, so I should not turn that into a win for paper. The trial did not show that paper was better, and failing to find a difference does not show two things equal. Its value is in how it asks. Hold the content and the time fixed, change one thing, and see what moves. ## What I would ask before setting a target The label “evidence-based” will not settle whether a target teaches. Federal law counts a significant result from a “correlational study with statistical controls for selection bias” as evidence, the third of its research tiers, and its school-improvement funds accept that tier [11]. So a usage comparison can lawfully carry the label. I would ask four things of any study behind a target. Was the program assigned, or the minutes, or neither? What did the comparison children do with the same time? Whose test measured the gain? Who ran the study, and who paid? A district can also test a schedule, within limits. It can assign two reasonable schedules to classrooms by lot for a term. Before it starts, it names a test the vendor did not write and the smallest gain worth the time. Gains of a tenth of a standard deviation are the size to expect, and the Russian trial needed 343 classes to measure one with a standard error of 0.039. So I reason that one district’s term will mostly show whether a schedule is workable. Districts that share one design and pool their results could show which schedule teaches more. The cost is the lottery, a second test and someone to analyze it, and the risk is that some classes get less of something that helps. The step can be undone, since a schedule can change next term. ## The case for the target The strongest case against me has three parts, and I accept each of them as far as it goes. Software has raised scores in trials that assigned it by lot. In Delhi a lottery offered children places in an after-school program that paired adaptive software with an instructor. After four and a half months the winners scored 0.37 standard deviations higher in math on tests the researchers designed [12]. A trial in Texas assigned teachers by lot to weekly coaching in how to run practice on Khan Academy. Across grades 3 to 8 together the effect was not significant. In grades 3 to 6, though, their students scored 0.12 to 0.17 standard deviations higher on the state test [13]. So pushing practice up can raise scores. What the trial assigned was a coach for the teacher, so it answers the adoption question for coaching. Usage reports describe, and the two described here state their own limits in plain words. A dashboard also tells a school what it could not otherwise see, which is whether the program it bought ever reached the children. And schools need targets, because a teacher cannot run a trial every Monday and needs a number to plan a week around. My objection begins where the number changes jobs. As a check that the schedule ran, a minutes target is sound. As evidence that the minutes taught, it is unproven, and as the goal itself it invites adults to extend a session to turn the dashboard green. A child who has mastered the lesson, or who is stuck and guessing, logs the same minute. That last worry is a guess about incentives, and no measurement of how often it happens was found for this essay. ## Check The strongest evidence against me comes from the same Delhi program. There the researchers made two estimates of what a day of attendance was worth. One came from the lottery. The other came from comparing children who attended more with children who attended less, adjusted for earlier scores. They could not tell the two apart [12]. In that program the usage comparison I distrust gave about the right answer, and the authors knew it only because they had a lottery to check it against. So usage comparisons can be right. In the heart trial, though, adjusting for 40 starting characteristics still left a gap of about 10 points [3]. A child’s fall score is one starting characteristic, and the i-Ready brief describes adjusting for that alone [1]. The essay leaves things out. It weighs test scores and ignores what software may give a teacher, such as time with a small group while the others practice. It ignores what a program costs to buy. The search for trials that assigned the amount was a search, not a systematic review, and its main result rests on one country and one program. Any school can run a check that could fail. Ask whoever produced the usage report to rerun it twice for the same two groups of children. The first run takes reading as the outcome, and the second takes the previous year’s growth, before the program began. If the children who met the target were already growing faster the year before, or gained as much in reading, the report measured the children. If both gaps are absent, one explanation is gone and my doubt about that report should shrink. I am confident that such a comparison cannot by itself show what requiring the target would do. It can be right, as in Delhi, but a school cannot know it is right without an assigned comparison to check it. I am much less sure about the right number of minutes for any program, and I do not think the evidence yet supports one. A trial would change my mind about a given target. Assign two schedules of one program by lot, in schools like the ones deciding, and measure on an independent test. If the longer schedule wins by a margin worth the hours, I would support it there. And if lotteries and usage comparisons of the same programs agreed as often as they did in Delhi, I would trust usage reports much further. ## Keep A met target shows who did the minutes, not what the minutes did. ## Sources 1. Curriculum Associates Research, June 2021. *Impact of i-Ready Personalized Instruction Time and Lesson Pass Rates on Student Learning Gains*, Research Report 2021-08. Written by the program’s publisher; the outcome is the publisher’s own test. [Summary page](https://www.curriculumassociates.com/research-and-efficacy/impact-time-lesson-pass-rates-student-learning-gains); [PDF](https://www.curriculumassociates.com/-/media/mainsite/files/i-ready/iready-instruction-dosage-study-2021.pdf) 2. Coronary Drug Project Research Group 1980, *New England Journal of Medicine* 303:1038–1041. Influence of adherence to treatment and response of cholesterol on mortality in the Coronary Drug Project. [PubMed](https://pubmed.ncbi.nlm.nih.gov/6999345/); [doi:10.1056/NEJM198010303031804](https://doi.org/10.1056/NEJM198010303031804) 3. Murray & Hernán 2016, *Clinical Trials* 13(4):372–378. Adherence adjustment in the Coronary Drug Project: a call for better per-protocol effect estimates in randomized trials. Reports the 1980 adjustment for 40 baseline characteristics. [doi:10.1177/1740774516634335](https://doi.org/10.1177/1740774516634335); [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC4942353/) 4. Murray & Hernán 2018, *Trials* 19:158. Improved adherence adjustment in the Coronary Drug Project: the placebo gap after accounting for events during follow-up. [PMC full text](https://pmc.ncbi.nlm.nih.gov/articles/PMC5836455/) 5. Grimaldi, Weatherholtz & Hill 2022, *Proceedings of the 15th International Conference on Educational Data Mining*. Estimating the causal effects of Khan Academy MAP Accelerator across demographic subgroups. The authors work at Khan Academy; observational. [ERIC full text](https://files.eric.ed.gov/fulltext/ED624104.pdf) 6. Pane, Doss, Todd & Seaman, RAND, June 2025. Efficacy of Zearn Math over two years in grades 3 to 5: an experiment in Texas. EdWorkingPaper 25-1211, funded by the Institute of Education Sciences; it also reports Zearn’s recommendation and Zearn’s own matched studies. [PDF](https://edworkingpapers.com/sites/default/files/ai25-1211.pdf) 7. Kraft 2020, *Educational Researcher* 49(4):241–253. Interpreting effect sizes of education interventions: benchmarks from 1,942 effects in 747 randomized trials. [doi:10.3102/0013189X20912798](https://doi.org/10.3102/0013189X20912798); [working-paper version, EdWorkingPaper 19-10](https://www.matthewakraft.com/s/interpreting-effect-sizes-of-educational-attainment.pdf) 8. Bettinger, Fairlie, Kapuza, Kardanova, Loyalka & Zakharov 2023, *Journal of Policy Analysis and Management* 42(2):552–570. Diminishing marginal returns to computer-assisted learning: workloads and usage records, pp. 554–555; Table 4, p. 561; acknowledgments, p. 567. [Author-hosted PDF](https://rfairlie.sites.ucsc.edu/files/2023/08/jpam-2023-marginal-returns-to-cal.pdf); [doi:10.1002/pam.22442](https://doi.org/10.1002/pam.22442) 9. Shah 2026, EdWorkingPaper 26-1379. The trade-off between quality and quantity: evidence from a field experiment on tutoring. One Indiana charter middle school, 343 students assigned by lot; tutoring twice a week in groups of two or three times a week in groups of three. [edworkingpapers.com](https://edworkingpapers.com/ai26-1379); [doi:10.26300/ma7z-9x69](https://doi.org/10.26300/ma7z-9x69) 10. Ma, Fairlie, Loyalka & Rozelle 2024, *Economic Development and Cultural Change* 72(4):1923–1962. Isolating the “tech” from EdTech: experimental evidence on computer-assisted learning in China. The figures here are from the working paper, Table 1. [NBER working paper 26953](https://www.nber.org/system/files/working_papers/w26953/w26953.pdf); [doi:10.1086/726064](https://doi.org/10.1086/726064) 11. 20 U.S.C. § 7801(21), the definition of “evidence-based,” with its stricter form for activities funded under section 6303. [law.cornell.edu](https://www.law.cornell.edu/uscode/text/20/7801) 12. Muralidharan, Singh & Ganimian 2019, *American Economic Review* 109(4):1426–1460. Disrupting education? Experimental evidence on technology-aided instruction in India: the lottery result, and the attendance estimates with their assumptions. [doi:10.1257/aer.20171112](https://doi.org/10.1257/aer.20171112); [author-hosted PDF](https://econweb.ucsd.edu/~kamurali/papers/Published%20Articles/Disrupting_education_AER.pdf) 13. Oreopoulos, Gibbs, Jensen & Price 2026, *Journal of Human Resources*, advance publication. Teaching teachers to use computer assisted learning effectively: experimental and quasi-experimental evidence. [doi:10.3368/jhr.1224-13977R2](https://doi.org/10.3368/jhr.1224-13977R2); [NBER working paper 32388](https://www.nber.org/papers/w32388) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/what-the-ai-earned/ Markdown: https://barani.org/writing/what-the-ai-earned/index.md Published: 2026-11-12 Modified: 2026-11-12 Form: Essay Topic: Healthcare Series: Hospital finance and care, part 3 of 4 ======================================================================== # How much of a paid claim did the AI earn? What a hospital must measure before it credits recovered revenue to artificial intelligence. Igor Barani · 12 November 2026 · 16 min read Healthcare · Essay · Hospital finance and care, part 3 of 4 · https://barani.org/writing/what-the-ai-earned/ An AI tool earned only the money the hospital would not have collected without it, and only a comparison fixed before the tool chooses its claims can measure that. Take one claim for $900. The insurer denies it for a wrong procedure code, the code is corrected, and the insurer pays. An AI tool touched the claim along the way, so its dashboard adds $900 to “recovered revenue.” On that one claim the tool may have earned $900, about a dollar, or nothing. It earned $900 if it found an error that the billing staff would never have found. It earned a week’s interest on $900, roughly a dollar, if the staff would have made the same correction seven days later. It earned nothing if it flagged a claim that someone was already fixing. The insurer’s payment looks the same in all three cases, because a payment records what happened and is silent on what would have happened otherwise. So the arrival of the money is an accounting event, and the tool’s share of it is a question of cause. Answering it takes two things. The first is a reliable history of each claim. The second is a comparison with similar claims that chance, and not the tool, set aside for the usual work. A hospital that pays on the dashboard’s total pays the tool for every claim it touched, including the ones that would have been paid anyway. As a physician I was trained not to credit a treatment because a patient improved after it. Some patients improve without it, so a treatment earns credit only for the difference between patients assigned to it and patients who were not. A claim paid after the tool touched it is that patient. ## Why the dashboard counts too much A list of the claims a tool touched describes its workload, and someone chose that list. The tool may choose the denials most likely to be overturned, or the staff may send it easy claims and keep the hard ones. Such claims are often paid with no new tool. In claims from 2019 that covered 30% of the Medicare Advantage market, insurers denied 17% of first submissions, and 57% of those denials were later overturned [1]. So the dashboard’s total can include money that the usual work would have collected. A fee on that total pays the vendor most for touching the claims most likely to be paid anyway. The question is pressing. In one seller’s survey of 204 finance and billing leaders, fielded in August 2026, 64% said they definitely or probably plan to invest in AI for claims and denials [2]. ## What the payment record can say The first need is a history of each claim, and the insurer’s payment file supplies only part of it. When an insurer finishes with a batch of claims, it sends the hospital an electronic file that lists each payment and each amount it held back. The file is the electronic remittance advice, called the 835 after the number of its transaction standard. An amount held back is an adjustment, and Medicare’s version of the file can adjust one line of a claim, a whole claim, or the provider’s account [3]. Each adjustment carries codes. A group code says who bears the unpaid amount, the provider or the patient, and a reason code gives the insurer’s reason [3]. The 835 is therefore good evidence of what the insurer decided. It does not say which person or program caused a correction, whether the correction was needed, or whether ordinary work would have brought the same payment. A payment in the file is also not yet cash that stays. An adjustment to the provider’s account, such as the recovery of an earlier overpayment, is usually tied to no claim in that file [3]. So when an insurer takes back a payment, the file may record the loss against the account, and a dashboard that adds up paid claims may never see the money leave. A seller of revenue-cycle software counted fewer denials on first submission in the first half of 2026 than a year earlier, 10.63% of claims against 12.00%. Yet cash came in no faster. Its researchers pointed to what insurers do after they pay, including taking payments back [4]. ## One set of claims gives two denial rates The history must also say which denial rate it reports, because a pilot can show a falling rate by changing the definition. Suppose a hospital sends 1,000 claims and follows every one to its end. The insurer denies 150 on first submission and later pays 60 of those after correction. These numbers are an example. The share denied at first is 150 in 1,000, or 15%. The share still denied at the end is 90 in 1,000, or 9%. A task force of the Healthcare Financial Management Association defines an initial denial rate that counts only the first denial of a claim and leaves resubmitted bills out [5]. KFF’s analysis of plans sold on HealthCare.gov does the reverse. A claim that was denied, resubmitted and paid is not counted as denied, and on that basis insurers denied 19% of in-network claims in 2024 [6]. The first rate asks how often the first attempt fails, and the second asks how much is lost in the end. The danger lies in moving from one to the other. A report that showed 15% last year and 9% this year shows a fall of 40% in which no claim fared any better. A fair count also waits. Claims denied last month have had little time to be appealed, so their final rate will look worse than that of claims a year old. Before a pilot starts, the hospital should therefore write down which claims are eligible, what the denominator is, how long each claim is followed, and how it counts partial payments and reversals. ## A correction can split one claim in two A correction can erase the history altogether. Return to the $900 claim, whose correction changed its procedure code. In Medicaid data, for instance, a corrected claim usually gets a control number of its own, so the claim’s number can change when it is sent again [7]. An analyst whose data lack the link between the two numbers may fall back on a key built from four fields. They are the patient, the provider, the date of service and the procedure code. The corrected claim now has a different key, so the analyst’s table holds two claims where the hospital sent one. > **Figure 1.** One $900 claim, denied under code A, corrected and paid under code B. A key that includes the code counts two unlinked claims, one denied and abandoned and one paid at first try, so the recovery vanishes. A schematic of an example claim, drawn for this essay. > > What the figure shows: Above, one claim as the hospital sent it: the first version, with procedure code A, is denied, then corrected, and the version with code B is paid 900 dollars, so one claim was recovered. Below, the same claim counted with a key that includes the procedure code: it becomes two records with no link between them, one denied and abandoned under code A and one paid at the first try under code B, and the recovery disappears. No dollar moved, but the history that would show a recovery is gone. KFF describes this effect in the federal data. When a provider files a new claim instead of resubmitting the original, “the original claim would be counted as denied, even if the new claim was ultimately paid” [6]. Suppose that in the 1,000-claim example, half of the 60 recoveries involved a change to a field in the key. Each of those 30 originals still looks denied and abandoned. The analyst then sees 1,030 claims and 30 recoveries among 150 denials, so the recovery rate reads 20% when the truth is 40%. The error can also differ between the two groups an evaluation compares. If the tool’s corrections change codes more often than the staff’s do, the key hides more of the tool’s recoveries, and the bias runs against the vendor. The repair belongs in the data, and the hospital claim already has a place for it. A replacement bill is marked as one by a digit in its bill type. A separate field carries the control number of the claim it replaces, and that number is the link [8]. A claim sent as new carries neither. ## The comparison must come before the tool chooses The second need is a comparison, and it must be fixed before the tool chooses, because the tool’s list of touched claims was selected. An insurer may also fix an error of its own across hundreds of claims, and each will sit on that list as a recovery. One health system told an HFMA reporter this year that it had been overturning 68% to 70% of the denials it appealed. With AI drafting part of some appeal letters, it put the rate at “probably 71% to 72% now” [9]. Taken at face value, about 70 of every 100 appeals were already being won. The tool could then claim one to four appeals in each 100, far fewer than 71 or 72, and a before-and-after count cannot show even that. The cleanest test borrows the clinical trial’s method. Before anyone works a batch of eligible denied claims, chance decides which go to AI-assisted work and which to usual work. Chance made the split, so the two groups hold the same mix of easy and hard claims, give or take luck. A later difference in collections can then be credited to the tool. This is random assignment. Three rules keep the test honest. The first is to count every claim in the group it was assigned to, including claims the tool never acted on. A tool that staff ignore has failed in a way the hospital pays for, and dropping its claims hides the failure. Trials call this rule intention to treat. The second is to assign at the right moment. A tool meant to prevent denials must be tested among claims not yet sent, because a study limited to claims still denied has selected on the outcome the tool should change. The third rule concerns the staff. If the same people work both groups, the time the tool saves on its claims can flow to the others. The usual-work group then does better than it would have, and the tool looks smaller than it is. The opposite error is also possible, because staff who know which claims are the tool’s may favor them. Where either is likely, chance should assign whole teams or sites, so that no one works both groups. In an example test, a hospital assigns 1,000 denied claims to each group at random and follows both for the same time. The AI-assisted group recovers 360 claims and the usual-work group recovers 300. To keep the arithmetic plain, every recovery pays $900 and none is reversed. The dashboard credits the tool with 360 × $900, or $324,000. The usual-work group collected $270,000, so the estimate of what the tool added is the difference, $54,000. The dashboard credits the tool with six times what it added. Suppose the tool’s fees and setup cost $20,000 and the staff time to review its output $10,000, so that $24,000 remains. > **Figure 2.** The example pilot in dollars, each bar from zero: the dashboard’s credit, the usual-work group’s collections, and the increment with its 95% range. The low end of the range, $17,000, falls short of the $30,000 of costs at the dotted line. Example numbers, drawn for this essay. > > What the figure shows: Three bars from zero for an example pilot of 1,000 denied claims in each group. The dashboard credits the tool with 324,000 dollars, the total collected in the AI-assisted group. The usual-work group collected 270,000 dollars, so the estimated increment is 54,000 dollars, one sixth of the dashboard’s credit, with a 95 percent range from 17,000 to 91,000 dollars. A dotted line marks costs of 30,000 dollars: the low end of the range falls short of them. Even $54,000 is an estimate. The tool’s lead is 6 points, 36% of claims recovered against 30%. With 1,000 claims in each group, the luck of the split leaves a 95% range of about 4 points either way, or $17,000 to $91,000. The range excludes zero, so at the usual 95% confidence the tool helps. Its low end does not cover the $30,000 of costs, so the pilot has not yet shown that the tool pays for itself. Real claims do not all pay $900. A few large ones may decide a pilot, so the range in dollars is wider than the range in claims. The chance split should therefore be made within bands of claim size, and the dollar difference reported with a range of its own. ## Predicting a denial is not preventing one A model that predicts which claims an insurer will deny has made the first of four links. Someone must then make a change that the record supports. The insurer must respond differently because of it. The gain must exceed the cost of the review. A high prediction score establishes the first link only. The authors of one recent model say so in their discussion. They trained it on 678,881 claim lines sent to a single insurer, of which 31.9% were denied. Whether showing those predictions to billing staff lowers the denial rate, they write, “cannot be answered retrospectively” [10]. A denial is also the insurer’s decision, and no finding that the bill was wrong. Federal inspectors sampled payment denials issued by 15 of the largest Medicare Advantage insurers in one week of June 2019. They estimated that 18% of the denied requests met Medicare’s coverage rules and the insurer’s own billing rules [11]. A model trained on denials can therefore learn the insurer’s mistakes, and then steer staff to change bills that were right. ## Three kinds of value, each counted once A tool can add value in three ways. The hospital may collect more in the end, collect the same money sooner, or spend less work collecting it. Suppose a tool brings forward $1,000,000 that would otherwise have arrived 30 days later. The hospital has not earned another $1,000,000. It has gained 30 days’ use of the money. If the hospital pays 8% a year to borrow, an example rate, that use is worth $1,000,000 × 0.08 × 30/365, or about $6,575. Counting the same $1,000,000 as both new revenue and faster cash counts it twice. A pilot can make the same mistake. At 60 days the AI-assisted group may lead only because its claims were paid sooner, so the difference should be reported at 60 days and again after every claim has settled. Speed is worth more than interest in some cases. Medicare gives a provider 120 days from receiving a claim decision to request the first appeal, and a late request can be dismissed [12]. Past a deadline of that kind, later can mean never. A tool that gets an appeal filed in time, where the staff would have missed the date, earns the whole $900. Work saved needs the same care. An hour that the tool frees becomes a cash saving only if the hospital pays for one hour less. If the hour goes to other claims and their revenue is counted, its wages cannot be counted as saved too. ## Pay for the difference A contract that pays the vendor a percentage of every dollar received on a touched claim rewards contact. A fee tied to the estimated increment pays for what the hospital wants. Such a fee needs the estimate kept current after the pilot. A small fixed share of eligible claims, chosen by chance and always worked the usual way, does that. The contract can then pay on each year’s estimate, or on the low end of its range. The contract should also name the record. Before signing, both sides should write down which claims enter the test, what the usual work is, how reversals and refunds are handled, and whether review time counts as a cost. Every event needs two dates, when it happened and when the hospital learned of it. Without the second, a March reversal of a January payment silently rewrites January, and a rerun report no longer matches the one the vendor was paid on. The hospital should be able to export the record and recompute the result, so that the evidence outlasts the contract. Money is not the only outcome to watch. A payment is no audit of the coding behind it, so a tool that raises collections by coding what the chart does not support has added risk. Fewer denials are also no gain if unsupported balances move to patients, and the group code shows when they do. An amount moved from the provider’s column to the patient’s has been billed to the patient, and nothing has been recovered. ## The case for trusting the dashboard A hospital losing money to denials cannot wait a year for a trial, and a vendor’s before-and-after numbers can be right. They are close to right when the tool works claims that no one worked before. If staff used to abandon every small denial, then nearly every dollar the tool recovers on small denials is new. An analysis of 234 million claims from 2019 notes that “many denied claims are never appealed” [13]. Where staff did work the claims, a fee on the dashboard’s total charges the hospital for money it would have collected anyway, $270,000 of the $324,000 in the example. The wait is shorter than it sounds. The tool starts on half of the eligible claims on the first day. The hospital therefore gives up only the gain on the other half, if there is one, and only until the test ends. In the example that is about $54,000 for each 1,000 claims held back. The step is reversible, since the hospital can then send all claims to the tool. Size is the harder objection. The example needs 2,000 eligible denials, and skewed dollars need more, so a small hospital must pool many months or rely on the baseline check below. A chance split of single claims is not the only sound design. A hospital can stagger the start across teams, sites or insurers, and compare each group’s change with the change in groups still waiting. That comparison of changes is called difference in differences. It rests on the assumption that the groups would have moved alike without the tool, which should be written down and tested against the months before the start. A plain before-and-after count remains the weakest design, because the insurer’s rules, the staff and the mix of claims all change while the tool is installed. ## Check The strongest point against this essay is that it has no evidence of the kind it asks for. No study found for this essay shows that dashboards overstate, and the sixfold gap comes from the example’s numbers. The published rates cut both ways. If 57% of denials were overturned in 2019, then 43% were not [1]. A tool that recovers claims from that 43% earns close to what its dashboard says. A search of PubMed and the open web on November 11 and 12, 2026, found no randomized or otherwise controlled test of AI in denial management with dollars collected as the outcome. The nearest studies concern other products. In a randomized crossover trial whose authors include the tool’s developers, an AI coding assistant sped 15 participants’ coding of long notes, with no payment data [14]. At one health system, a difference in differences found that billed work rose more among physicians who chose an AI scribe than among those who did not. The share of claims with a denial did not differ [15]. Such a search is no systematic review, and vendors may hold trials they have not published. The essay ignores three things. Insurers can change how they decide, so a tool’s effect this year may not hold next year. The usual work is treated as fixed, though staff improve. The analyst’s time is not priced. A hospital can run a check that could prove the essay wrong for its own case. Before looking at results, list the kinds of denial the tool now works, each defined by insurer and reason code. For those kinds, compute the share of denied claims that were paid within a fixed window, say 180 days, in the year before the tool and in its first year. If the earlier share is half the present one, the dashboard credits the tool with twice what it added. If the earlier share is under about a tenth of the present one, the dashboard’s total is within about a tenth of what the tool added. There the essay overstates the problem. The check is itself a before-and-after count, so its answer is a first estimate, not a verdict. I am confident of the logic, because a payment after a touch cannot show cause and each kind of value can be counted only once. I am much less sure of the size. I do not know how far dashboards overstate at a typical hospital, and it may be little where the tool works claims that were being abandoned. A published trial, or a hospital’s own baseline, showing touched-claim totals within a tenth of the measured increment would change what I advise about contracts. It would not change the question. ## Keep Pay the tool for the difference from claims it never chose, not for every claim it touched. ## Sources 1. Vabson, Hicks & Chernew, *Health Affairs* 44(6):702–706, June 2025. Medicare Advantage denies 17 percent of initial claims; most denials are reversed, but provider payouts dip 7 percent. Claims from plans covering 30% of the Medicare Advantage market in 2019. [doi.org](https://doi.org/10.1377/hlthaff.2024.01485) 2. Experian Health, October 28, 2026. State of Claims 2026: a survey of 204 healthcare finance and billing leaders, fielded August 5–23, 2026. Experian Health sells claims software. [PDF](https://www.experian.com/content/dam/noindex/na/us/healthcare/state-of-claims-2026.pdf) 3. Centers for Medicare & Medicaid Services. Health Care Payment and Remittance Advice: the 835, adjustments at line, claim and provider level, group codes, claim adjustment reason codes and provider-level balance codes. Retrieved November 12, 2026. [cms.gov](https://www.cms.gov/medicare/coding-billing/electronic-billing/health-care-payment-remittance-advice) 4. TechTarget RevCycle Management, September 23, 2026. Insurers deny less just to claw back more later, analysis finds. Reports Kodiak Solutions’ revenue-cycle data: initial denial rate 10.63% in the first half of 2026 against 12.00% a year earlier, and the researchers’ account of post-payment activity. TechTarget describes the data as drawn from 2,400 hospitals. Kodiak sells revenue-cycle software and services. [techtarget.com](https://www.techtarget.com/revcyclemanagement/news/366651012/Insurers-deny-less-just-to-claw-back-more-later-analysis) 5. HFMA Claim Integrity Task Force, 2022. Standardizing denial metrics for revenue cycle benchmarking and process improvement: the initial denial rate counts “only the first denial that was made for a given claim” and excludes rebills by bill type. [hfma.org](https://www.hfma.org/guidance/standardizing-denial-metrics-revenue-cycle-benchmarking-process-improvement/) 6. Long, Lo & Pestaina, KFF, March 24, 2026. Claims denials and appeals in ACA Marketplace plans in 2024: an analysis of insurers’ reports to CMS, with its treatment of claims denied and later paid. [kff.org](https://www.kff.org/patient-consumer-protections/claims-denials-and-appeals-in-aca-marketplace-plans-in-2024/) 7. Centers for Medicare & Medicaid Services. PERM SC fast facts for claim adjustments: “ICNs of adjusted claims usually end in a number other than one and are tied to the original claim.” Describes state Medicaid claims data. Retrieved November 12, 2026. [cms.gov](https://www.cms.gov/files/document/claim-adjustment-fast-facts.pdf) 8. Mississippi Division of Medicaid, 2022. UB-04 Claim Form Version CMS-1450, section 5.8, UB-04 claim form instructions for institutional claims (Gainwell Technologies, version 1.0): bill-type frequency codes for interim, replacement and void claims, and Form Locator 64, “Enter the original claim ICN that is being requested to be replaced/adjusted or voided.” One state program’s instructions for the national hospital claim form. [PDF](https://medicaid.ms.gov/wp-content/uploads/2022/10/Section-5_UB40-Claim-Form-Version-CMS-1450.pdf) 9. Williams, Healthcare Financial Management Association, January 30, 2026, updated February 6, 2026. Battle of the bots intensifies over denials: a health system’s own account of its appeal overturn rate, with no comparison group described. [hfma.org](https://www.hfma.org/revenue-cycle/denials-management/battle-of-the-bots-intensifies-over-denials/) 10. Jadoon et al., *Frontiers in Health Services* 6:1927175, October 14, 2026. A machine learning framework for predicting healthcare insurance claim denials: retrospective, 678,881 claim lines, one insurer. [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC13713848) 11. U.S. Department of Health and Human Services, Office of Inspector General, April 2022. Some Medicare Advantage organization denials of prior authorization requests raise concerns about beneficiary access to medically necessary care, OEI-09-18-00260: a stratified random sample of 250 prior authorization denials and 250 payment denials, June 1–7, 2019. [PDF](https://oig.hhs.gov/oei/reports/OEI-09-18-00260.pdf) 12. Centers for Medicare & Medicaid Services. First level of appeal: redetermination by a Medicare contractor. A party “has 120 days from the date of receipt of the initial claim determination to file a redetermination request.” Retrieved November 12, 2026. [cms.gov](https://www.cms.gov/medicare/appeals-grievances/fee-for-service/first-level-appeal-redetermination-medicare-contractor) 13. Weinreb & Landon, *Health Affairs Scholar* 4(8):qxag206, August 2026. Variation in medical claim denials: safety-net providers are hardest hit. 234.2 million claims from 2019 across Medicare Advantage, Medicaid and commercial plans. [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC13492083) 14. Chomutare et al., *Journal of Medical Internet Research* 27:e71904, July 3, 2025. Artificial intelligence to improve clinical coding practice in Scandinavia: crossover randomized controlled trial. 15 participants and 300 coded notes analyzed; the tool’s developers are among the authors. [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC12244276) 15. Holmgren et al., *JAMA Network Open* 9(1):e2553233, 2026. Ambient artificial intelligence scribes and physician financial productivity: a cohort study at one health system, 1,565 physicians, January 2023 to April 2025, difference in differences. [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC12789954/) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/budget-neutral-340b/ Markdown: https://barani.org/writing/budget-neutral-340b/index.md Published: 2026-11-11 Modified: 2026-11-12 Form: Essay Topic: Healthcare Series: Hospital finance and care, part 2 of 4 ======================================================================== # What does a budget-neutral cut to 340B drug payments do to cancer care? What Medicare's 2027 rule moves between cancer services, and what it leaves alone. Igor Barani · 11 November 2026 · revised 12 November 2026 · 16 min read Healthcare · Essay · Hospital finance and care, part 2 of 4 · https://barani.org/writing/budget-neutral-340b/ By itself, it changes no treatment. A patient who receives her infusion on Monday, December 28, 2026, and again a week later, after the cut begins on January 1, gets the same drug from the same nurse. What changes is the payment. If her hospital bought the drug at the discount that a federal program called 340B requires of manufacturers, Medicare pays the hospital about a third less for the second dose. A few kinds of hospital and of drug are exempt [1]. If she also has radiation there, the rule that cuts the drug payment raises what Medicare pays for each treatment by 7.5%, because Medicare must keep its outpatient spending level [1, 2]. Her care is the same in both departments, but the cancer program's books now show infusion's margin falling and radiation's rising. CMS, the agency that runs Medicare, expects most 340B hospitals to lose by the trade [1]. A hospital that loses may have to cut a service because it can no longer pay for it. A second harm needs no loss at all. Leaders may read that price change as a change in how well each service is run. I think that misreading is the risk a hospital should act on first. ## What happens to one dose Take one dose of a cancer drug, and suppose the manufacturer nets $1,000 for it on average across its buyers. That average, called the drug's average sales price, leaves out the 340B sales and certain federal sales that the law excludes [1, 3]. The manufacturer reports it to Medicare each quarter. The $1,000 is an example, and the essay follows it throughout. When a hospital gives the dose in its outpatient department, Medicare's usual payment is the average sales price plus 6%, or $1,060. Some hospitals buy the dose for far less than $1,000. Hospitals that serve many low-income patients can join the 340B program, named for its section of the Public Health Service Act. The program obliges manufacturers to sell them outpatient drugs at a discount [2]. In 2026 Medicare pays such a hospital the same $1,060. From January 1, 2027, it will pay the average sales price minus 28%, or $720 [1]. The payment falls by $340, which is 32.1% of what Medicare paid before. The fall exceeds 28% because the old payment sat 6% above the price and the new one sits 28% below it, and 34 points out of 106 is 32.1%. For each drug the program sets a ceiling price, the most a manufacturer may charge a 340B hospital. Across drugs the ceiling runs 28% below the average sales price, so 28% is the smallest discount a hospital gets [1]. Many hospitals negotiate under the ceiling, and those CMS surveyed this year reported paying 33.4% below the average sales price [1]. CMS proposed a rate of 33.4% in July and finalized 28%. It called the ceiling data objective and conservative, and it cited commenters' warnings that the deeper cut would limit patients' access to the drugs [1]. At the proposed 33.4%, Medicare would have paid $666 for the example dose. A hospital that bought it at the average ceiling, $720, would then have been paid less than it spent. ## What the hospital keeps A hospital watches the gap between what Medicare pays for a dose and what the dose cost. Call that gap the spread. Suppose the hospital bought the $1,000 dose for $666, the average price hospitals reported in CMS's survey. It keeps $394 in 2026 and $54 in 2027. The payment falls by 32%, but the spread falls by 86%, because the cost does not move and the whole $340 comes out of the part the hospital kept. A hospital that pays the full ceiling price, $720, keeps nothing on the drug in 2027. > **Figure 1.** One $1,000 dose, as Medicare pays for it and as the hospital buys it. The payment falls from $1,060 to $720, by 32%, but the part the hospital keeps at the survey’s average cost of $666 falls from $394 to $54, by 86%. An example drawn for this essay from the rule’s rate and CMS’s survey average. > > What the figure shows: Two bars for one dose of a drug whose average sales price is one thousand dollars. In 2026 Medicare pays 1,060 dollars; at the survey average cost of 666 dollars the hospital keeps 394 dollars. In 2027 Medicare pays 720 dollars and the hospital keeps 54 dollars. The payment falls by 32 percent, and the part kept falls by 86 percent. The spread is not the margin on the visit. Medicare also pays to administer the drug, and the hospital pays for pharmacy preparation, nursing time, storage, waste and the care it gives without charge. A cancer service therefore has two questions to answer. Does the payment cover the medicine? Does all the revenue from the visit cover what the visit uses? ## Where the money goes Medicare pays hospitals for outpatient care from one schedule of rates set in advance, the outpatient prospective payment system, or OPPS, which this essay calls the outpatient system [4]. The law requires CMS to keep total spending in the outpatient system unchanged when it changes relative payments [2]. That requirement is called budget neutrality. The cut takes $4.65 billion out of drug payments in 2027, counting the share that patients pay themselves. So the same $4.65 billion goes back out as a 7.5% raise on non-drug services at every hospital the system pays [1]. CMS's headline figure of $5.45 billion also counts $0.8 billion cut at off-campus departments paid under other rules, which the raise does not return [1]. For-profit hospitals cannot take part in 340B, so they collect the raise and pay none of the cut [5]. A budget will show a rise larger than 7.5%. Medicare sets each service's payment by multiplying a weight, which ranks the service's typical cost, by one dollar amount called the conversion factor [4]. The rule raises that amount by 9.4%, from $91.415 to $99.970. Besides the 7.5%, the rise carries the yearly update of 2.3%, less about half a point of technical adjustments, the largest for local wages [1]. That update would have come without the 340B rule, so it belongs in the budget and out of any judgment of the rule. ## What the cut spares The cut reaches only drugs that Medicare pays for one by one. Medicare folds cheaper drugs, like most supplies, into the rate for the visit that uses them. Most drugs that cost more than $135 a day in 2027 get their own payment [1, 4]. Only those drugs take the cut, and radioactive drugs, called radiopharmaceuticals, are among them. Vaccines, non-opioid pain drugs and drugs with pass-through status, a temporary status that many newly approved drugs hold, keep the old rate [1]. Three kinds of hospital keep the old drug rate: rural sole community hospitals, children's hospitals, and the 11 cancer hospitals exempt from Medicare's inpatient payment system [1, 6]. They still collect the raise, which is why CMS estimates a 5.3% gain for rural sole community hospitals [2]. A cancer department inside a general hospital is not one of them. Medicare does not pay critical access hospitals through the outpatient system, so neither the cut nor the raise reaches them [2]. ## One ratio decides who loses Every hospital in the outpatient system gets the raise, but only some take the cut, so a hospital's result depends on how much of each kind of payment it has. Hold drug prices, volumes and billing fixed, and only two numbers in its books matter. One is its traditional Medicare outpatient payment for the affected drugs, which falls by 32.1%. The other is its traditional Medicare outpatient payment for the services that receive the raise, which rises by 7.5%. Call them the two streams. The hospital loses 32.1 cents on each drug dollar and gains 7.5 cents on each service dollar. It therefore needs 32.1 ÷ 7.5, or 4.28, service dollars to make up for one drug dollar. One drug dollar among 5.28 is 19%, so the two effects cancel when affected drugs make up 19% of the two streams together. Below that share the hospital gains, and above it the hospital loses. A hospital that divides by total revenue instead will report a share far below 19% and may conclude that it gains when it loses. Consider three example hospitals, each with $100 million a year in the two streams. Hospital A has $10 million in affected drugs and $90 million in services, so it loses $3.21 million on the drugs, gains $6.75 million on the services and ends $3.5 million ahead. Hospital B, with $20 million in affected drugs, loses $0.4 million. Hospital C, with $30 million, loses $4.4 million. The same national rule produces opposite results because the hospitals sell different mixes of care. Many hospitals also owe a repayment, which lowers their break-even share. From 2018 to 2022, Medicare cut payment for 340B drugs and raised everything else to match, until the Supreme Court struck the cut down. CMS then paid the 340B hospitals what the cut had withheld [1], which left the earlier raise with no cut to pay for it. So CMS began taking back $7.8 billion from the hospitals that had received that raise, by trimming each year's rate for non-drug services [5]. The trim applies to hospitals enrolled in Medicare before 2018 [1]. Such a hospital does not see the full 7.5%. The same rule quickens the trim, from half a point in 2026 to two points in 2027 [1]. CMS takes it out of the yearly update that Medicare gives the service rate for rising costs [4]. A repaying hospital is therefore paid about 99.5% of the full service rate in 2026 and about 98% in 2027. The 7.5% raise turns that 98% into about 105.4%. From 99.5 to 105.4 is a rise of 5.9%, and that is the raise this hospital should count. A smaller raise makes up for fewer drug dollars, so the break-even share falls from 19% to about 15.5%. Hospital A's gain becomes $2.1 million, and Hospitals B and C lose $1.7 million and $5.5 million. > **Figure 2.** How the rule changes a hospital’s two streams, taken together, by the share of them that is affected drugs. The solid line is the drug cut and the raise alone, which break even at about 19%. The dotted line adds the faster repayment, which moves break-even to about 15.5%. Drawn for this essay from the rates in the text, with A, B and C for the three example hospitals. > > What the figure shows: Two falling lines show the net effect of the 2027 340B rule on a hospital, as a percent of its affected drug payments plus the non-drug payments that get the raise, against the drug share of those two. The solid line, for the drug cut and the 7.5 percent raise alone, starts at plus 7.5 percent and crosses zero at about 19 percent. The dotted line, which adds the faster repayment, starts at plus 5.9 percent and crosses zero at about 15.5 percent. Dots mark example hospitals at shares of 10, 20 and 30 percent, which change by plus 3.5, minus 0.4 and minus 4.4 percent. CMS's own estimates of the 340B policy, by type of hospital, fit the ratio. The hospitals with the largest shares of low-income patients, which KFF calls safety-net hospitals, lose 5.2% of their payments from the outpatient system. Urban hospitals of 500 beds or more lose 4.8%, and major teaching hospitals lose 3.9% [2]. For-profit hospitals gain 6.8%, which is what the 7.5% raise would give on roughly nine tenths of their payments. ## A cancer program holds both streams Cancer drugs made up 41% of all 340B purchases in 2021, the largest share of any class [7]. A cancer program in a 340B hospital therefore holds both streams. Its infusion center gives the drugs that take the cut. Its radiation department gives treatments that receive the raise. Radioactive drugs given as treatment are the exception: they take the cut when bought through 340B and paid for separately [1]. A cancer program also runs services that no claim pays for, such as navigation, social work, toxicity management and the coordination that carries a patient between treatments. Whether a hospital paid for them out of drug margins is a fact in its own books. If it did, that money is now smaller, so each service needs its own line showing what it costs and how much of that cost no payment covers. A board can fund a named shortfall, but it cannot weigh a margin that paid for services without saying so. ## A payment change is not a performance change I am a radiation oncologist, and this rule flatters my own specialty. In a 340B hospital it will make hospital-based radiation look better and infusion look worse in the same cancer program, in the same month, for the same patients. An infusion service can look less productive while it treats as many patients with the same nurses, and radiation can look more profitable without one better outcome. That matters because margins steer decisions. A compensation plan that rewards a margin increase will reward a price change that no clinician made. A capital committee that reads a margin decline as underperformance may take chairs, staff or hours from a service the community still needs. I describe the first path as plausible. The second I have seen, though only indirectly, and I have not measured it. I would ask a cancer program to read two financial views side by side. The first uses the 2027 prices and shows the money available to operate. The second holds 2026 prices fixed, so that a change in volume, cost or mix shows up alone. Outcomes and access measures belong beside both. For a patient who needs chemotherapy and radiation, the unit of judgment is the whole course of treatment. A gain in one department's ledger does not pay for a loss in another's while their budgets, staff and access targets stay separate. ## The hospital may not yet know which doses take the cut Both views rest on knowing which doses were bought through 340B, and a hospital's enrollment in the program does not say. Many 340B hospitals keep one physical stock of a drug and give every dose from it. Software then sorts the doses by whether each patient was eligible for the 340B price, and the hospital buys each replacement on the matching account [8, 9]. So the claim can go out before that sorting is final [1]. From January 1, the claim must say how the drug was bought. Every separately paid drug line carries a two-letter code. JG marks a 340B drug that takes the cut, and the new ND a drug bought outside 340B. TB marks a drug reported for information only, such as a 340B drug at an exempt hospital [1]. A dose matched to the wrong purchase, or matched too early, therefore misstates both the payment and the hospital's exposure. The remedy is a reconciliation that starts with the drugs that carry the most money, matches each dose to its purchase record and sends every doubtful line to a queue. Three counts show its state: the unmatched dollars, the lines later reclassified, and the days from dose to final classification. An unresolved line waits in the queue and is never counted as bought outside 340B by default. ## What I would put in front of a board 1. The exposure: affected drug payments, the payments that get the raise, the share of the two, and each step from that share to the budget. For infusion, add the whole visit's revenue against its whole cost. 2. Its reliability: how many drug dollars the reconciliation matched, how many remain unresolved, and how much rests on claims rather than purchase records. 3. The services at stake: what the affected margin paid for, what each costs, and the other ways to fund it. 4. The evidence after January: realized payment and cost, delays to treatment, patients referred elsewhere, against a baseline set before the rule. ## The strongest case against me The strongest argument for the rule is about incentives. Medicare's payment is a percentage of the drug's price, and so is the 340B discount, so the hospital's margin grows with the price of the drug and the number of doses. The Congressional Budget Office judged that 340B encourages behavior that raises federal spending, such as opening off-site clinics, though it could not measure how much [10, 11, 12]. That judgment is not proof that a given treatment was unneeded or that a given hospital's margin was too large. The incentive argument cuts against me. If drug margins drove care that patients did not need, then part of the fall in infusion margins is a correction. A correction would explain why infusion margins fall, but not why radiation's rise. A patient's own share of an outpatient payment, called coinsurance, is usually 20% [4], so on the example dose it falls from $212 to $144. CMS expects a person who receives a qualifying drug to save $900 on average in 2027 [13]. The raise for other services increases what patients pay for those [2], and the patient who saves and the patient who pays more need not be the same person. One supporter of a deeper cut says the rule stops short, because a payment set at the ceiling still leaves the average surveyed hospital a spread [14]. The Community Oncology Alliance, which speaks for independent oncology practices, welcomed the rule as a correction of excessive payments for 340B drugs [15]. The American Hospital Association calls it an assault on 340B [6]. The other case against me is the plain loss. Hospital C ends $5.5 million behind, and a hospital in that position may cut a service because it cannot pay for it, whether or not anyone misreads a margin. A constant-price view does not restore that money. I put the misreading first because it is the one a hospital can avoid. I have also seen it, though only indirectly, and the pressure it brought had the reverse impact, as the Check describes. ## Check A test is open to any hospital. Reprice the 2026 outpatient claims at 2027 rates with every volume fixed, and compare the infusion and radiation margins before and after. If the gap between them moves by less than the normal year-to-year swing, the rule is too small to mislead anyone, and this essay overstates its importance. A correction and a misreading can be told apart after January. A correction shows up as fewer doses of drugs that guidelines do not support. A misreading shows up as fewer chairs, nurses or hours while waits and referrals elsewhere grow. The nearest test of my concern is the last cut. From 2018 to 2022, Medicare paid 340B drugs at the average sales price minus 22.5%. That cut was smaller, a fall of 27% against 32.1%, and it was later repaid, so it tests this one loosely. Two studies of traditional Medicare claims show what hospitals did in those years. In the first, drug claims at the older off-campus departments of 340B hospitals rose from 3.4% to 4.9% of outpatient services between 2016 and 2022. At eligible hospitals outside the program the share stayed near 2% [16]. In the second, hospitals that joined 340B in 2019, while the cut was in force, added drug administration at new off-campus sites [17]. The first study is descriptive, and the second compares 122 newly joined hospitals with 779 others. Neither measures a response to the cut. Still, drug delivery kept growing while the cut was in force. That is weak evidence against my fear of withdrawn capacity and none for it. If the pattern repeats, the risk moves from fewer chairs to more doses, and a constant-price view shows that too. In my experience, finance teams model new rates every year, so they may already separate price from performance, which weighs against me. Yet in the cases I have seen, though only indirectly, leaders answered a payment cut by pressing the affected departments to improve efficiency and cut staff. That pressure had the reverse impact. This is my experience, not a measurement. The 19% line leaves things out. It holds the average sales price fixed, though that price moves every quarter. It ignores the yearly reweighting of each service, the adjustment for local wages and the extra payments for unusually costly cases [4]. It ignores the same rule's new cuts to imaging at older off-campus sites [1]. It counts only traditional Medicare, not commercial contracts or the Medicare Advantage plans that private companies run [18]. The law is unsettled. The 2022 decision held that Medicare may not set different drug payments for groups of hospitals without first surveying what they paid [19]. CMS has now surveyed them. The American Hospital Association told CMS that its proposals had serious legal defects [20], and I found no ruling on the final rule. I am confident of the direction. Hospitals whose affected drug payments are large against their other Medicare outpatient payments will lose, and cancer programs will carry much of that loss. I am much less sure of any one hospital's dollars, which need its own reconciled data. I am least sure how often the misreading happens, which is my judgment of how margins are used, not a measurement. Evidence that the 2027 change moves volume rather than capital, as the last cut appears to have done, would change what I fear, but not what I advise. ## Keep The rule changes what a dose is paid, not what it does, so judge a cancer program at constant prices. ## Sources 1. CMS, CY 2027 OPPS final rule with comment period (CMS-1850-FC), 91 Fed. Reg. 70284–70762, November 6, 2026: the 28% rate and its reasons, the exclusion of 340B prices from the average sales price, the $135 threshold, the 11 cancer hospitals, the $4.65 billion and $5.45 billion, the 7.5% raise, the conversion factor of $99.970, the trim from 0.5% to 2%, radiopharmaceuticals, and modifiers JG, TB and ND. [Federal Register](https://www.federalregister.gov/documents/2026/11/06/2026-22739/medicare-program-hospital-outpatient-prospective-payment-and-ambulatory-surgical-center-payment) 2. Hulver, Levinson, Cubanski & Freed, KFF, November 6, 2026. Medicare cuts to 340B drug payments will hit safety-net hospitals while benefiting for-profit hospitals: budget neutrality, critical access hospitals, cost sharing and CMS's estimates by hospital type. [kff.org](https://www.kff.org/health-costs/medicares-proposed-cut-to-340b-drug-payments-would-hit-safety-net-hospitals-while-benefiting-for-profit-hospitals/) 3. Social Security Act § 1847A(c)(2), 42 U.S.C. 1395w-3a(c)(2), and § 1927(c)(1)(C)(i), 42 U.S.C. 1396r-8(c)(1)(C)(i): the sales the average sales price leaves out. [law.cornell.edu](https://www.law.cornell.edu/uscode/text/42/1395w-3a) 4. Medicare Payment Advisory Commission, October 2026. Payment basics: outpatient hospital services payment system. How a rate is set from a weight and the conversion factor, which drugs are paid separately, the wage and outlier adjustments, and the 20% copayment. [PDF](https://www.medpac.gov/wp-content/uploads/2026/10/MedPAC_Payment_Basics_26_OPD_FINAL_SEC.pdf) 5. Healthcare Financial Management Association, November 2026. 2027 Medicare OPPS final rule cuts 340B payments, expands site-neutral rates: the $7.8 billion repayment, its history, and for-profit hospitals outside 340B. [hfma.org](https://www.hfma.org/payment-reimbursement-and-managed-care/2027-medicare-opps-final-rule-340b-payments/) 6. American Hospital Association, November 4, 2026. CMS finalizes increases to Medicare hospital outpatient department payment rates, changes to site-neutral and 340B for 2027: the exemptions and the association's response. [aha.org](https://www.aha.org/news/headline/2026-11-04-cms-finalizes-increases-medicare-hospital-outpatient-department-payment-rates-changes-site-neutral-and) 7. Sen, Congressional Budget Office, testimony to the Senate HELP Committee, October 23, 2025. Growth in the 340B Drug Pricing Program and its implications for the federal budget: cancer drugs 41% of 340B purchases in 2021. [PDF](https://www.help.senate.gov/download/sen-testimonypdf) 8. American Hospital Association, August 7, 2026, revised October 2026. Fact sheet on the 340B rebate model, including the replenishment model hospitals use for inventory. [aha.org](https://www.aha.org/fact-sheets/2026-08-07-fact-sheet-340b-rebate-model-would-undermine-access-care) 9. Crowe LLP. 340B Program self-monitoring: purchasing and inventory management. Inventories "physically comingled but kept virtually separate using a replenishment model," with software for eligibility and accumulation. [crowe.com](https://www.crowe.com/insights/healthcare-connection/340b-program-self-monitoring-purchasing-and-inventory-management) 10. Congressional Budget Office, September 9, 2025. *Growth in the 340B Drug Pricing Program*. [cbo.gov](https://www.cbo.gov/publication/60661) 11. Applied Policy, September 2025. Congressional Budget Office publishes report on 340B program growth, 2010–2021, quoting CBO on behaviors that tend to increase federal spending. [appliedpolicy.com](https://www.appliedpolicy.com/?p=10027) 12. Healthcare Financial Management Association, September 19, 2025. Proposed 340B rebate model generates an array of concerns among providers, reporting the factors CBO could not quantify. [hfma.org](https://www.hfma.org/accounting-and-financial-reporting/proposed-340b-rebate-model-generates-an-array-of-concerns-among-providers/) 13. CMS, November 4, 2026. Fact sheet, CY 2027 OPPS and ASC final rule: the $900 average saving. [cms.gov](https://www.cms.gov/newsroom/fact-sheets/calendar-year-2027-hospital-outpatient-prospective-payment-system-opps-ambulatory-surgical-center-0) 14. Hammond, Paragon Health Institute, November 2026. CMS finalizes 2027 outpatient and physician payment rules: the ceiling as the minimum discount. Policy advocacy that supported the deeper cut. [paragoninstitute.org](https://paragoninstitute.org/paragon-prognosis/cms-finalizes-2027-outpatient-and-physician-payment-rules/) 15. Targeted Oncology, November 6, 2026. The Community Oncology Alliance's reaction to the 2027 final rules. [targetedonc.com](https://www.targetedonc.com/view/coa-reaction-cms-2027-final-rules-340b-physician-payment) 16. Odouard, Thome & Buntin, *Medical Care Research and Review* 83(6):563–572, June 2026. Changes in service delivery following site-neutral payment reforms by hospital 340B status: descriptive, 2,550 hospitals, traditional Medicare claims 2016–2022. [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC13740177) 17. Thome, Odouard & Buntin, *JAMA Health Forum* 7(10):e263977, 2026. Hospital 340B participation, site-dependent payment, and vertical integration: difference in differences, 122 new 340B hospitals against 779 others, traditional Medicare claims 2016–2019. [PMC](https://pmc.ncbi.nlm.nih.gov/articles/PMC13723016) 18. Medicare.gov. Your coverage options: a Medicare Advantage plan is "a Medicare-approved plan from a private company." [medicare.gov](https://www.medicare.gov/basics/get-started-with-medicare/get-more-coverage/your-coverage-options) 19. *American Hospital Association v. Becerra*, 596 U.S. 724 (2022), decided June 15, 2022, unanimous. [Supreme Court](https://www.supremecourt.gov/opinions/21pdf/596us2r46_bqm1.pdf) 20. American Hospital Association, August 26, 2026. Comments urging CMS not to finalize the 340B cut or the faster repayment, citing legal defects. [aha.org](https://www.aha.org/letterscomments/2026-08-26-aha-urges-cms-not-finalize-opps-proposals-accelerate-clawback-and-cut-340b-reimbursements) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/first-point-of-aries/ Markdown: https://barani.org/writing/first-point-of-aries/index.md Published: 2026-11-05 Modified: 2026-11-05 Form: Note Topic: Origin Series: Type, figures and tools, part 1 of 5 ======================================================================== # Why is the first point of Aries still named for Aries? The Sun’s March crossing point left the Ram two thousand years ago. Igor Barani · 5 November 2026 · 8 min read Origin · Note · Type, figures and tools, part 1 of 5 · https://barani.org/writing/first-point-of-aries/ On the day of the March equinox the Sun stands in front of the stars of Pisces, as it has every March for two thousand years. Yet the point it occupies that day is still called the first point of Aries. The name stays because it stopped describing the stars behind the point and came to mean the point’s job. That job is to mark where the Sun crosses the sky’s equator heading north, the zero from which the sky’s longitude is counted. The stars behind the zero have changed, but the job has not. ## Where the sky is measured from A position on Earth needs a longitude, and a longitude has to be counted from an agreed zero. A position in the sky needs a zero too, and the sky supplies one. Picture the sky as a great sphere around Earth, with two circles drawn on it. The first is the celestial equator, which is Earth’s equator carried out onto the sphere. The second is the path the Sun appears to follow against the stars over a year, called the ecliptic. The two circles are tilted about 23.4 degrees to each other [1], and two tilted circles on one sphere cross at two opposite points. At one of those crossings the Sun passes from south of the equator to north of it, and that moment is the March equinox. The crossing itself is the first point of Aries. Right ascension, which is the sky’s longitude, is counted from it [2], and charts mark it with the Ram’s sign, ♈ [2]. ## The view that gave the name The Greek astronomer Hipparchus defined the point in about 130 BC, when it lay at the western edge of the constellation Aries, the Ram [2]. The Sun travels eastward along its path [1], so the point was the first of the Ram that the Sun reached each year. The name described the view. > **Figure 1.** The stars along the Sun’s path, from Taurus to Aquarius, with a time line beneath on the same scale. The March equinox point drifts west one degree in about 72 years, so each place on the strip is also a date. The ring marks the point, and west is to the right, so it moves right as the years pass. The point entered Aries in 1866 BC, was defined there in 130 BC, left for Pisces in 68 BC and reaches Aquarius in 2597. Move the slider to choose a year. Dates from March equinox, *Wikipedia* [3]. Rate from Axial precession, *Wikipedia* [1]. Stars from the Hipparcos-based catalog in d3-celestial. > > What the figure shows: A strip of the stars along the ecliptic, from Taurus on the left through Aries and Pisces to Aquarius on the right, with a time line beneath it on the same scale. The March equinox point drifts to the right, westward, about one degree in 72 years. It crossed from Taurus into Aries in 1866 BC, was defined by Hipparchus in 130 BC, left Aries for Pisces in 68 BC, stands in Pisces today and reaches Aquarius in 2597. Hipparchus also found the first sign that the view would not last. The equinox is not marked in the sky, so he needed the Moon as a reference [1]. During a lunar eclipse the Moon stands exactly opposite the Sun, and the Sun’s place on a given date could be calculated. So an eclipsed Moon gave him a point whose distance from the equinox he knew. He is thought to have measured how far the star Spica stood from it [1]. Then he compared his result with measurements by Timocharis, who lived from about 320 to 260 BC. Spica’s place, counted from the equinox, had shifted by about 2 degrees [1]. Either the star had moved or the zero had. Hipparchus noticed the same motion in other stars and concluded that the equinoxes themselves were moving [1]. The reason lies in the shape of the Earth. ## Why the point slides Earth bulges at the equator, and its axis is tilted, so the bulge lies at an angle to the Sun’s path. The Sun and the Moon pull harder on the half of the bulge nearer to them, because gravity weakens with distance [1]. That uneven pull drags the tilted bulge toward the plane of the Sun’s path, which would stand the axis upright. On a perfect sphere the pulls would balance, and nothing would turn [1]. Because Earth is spinning, the pull does not straighten the axis. The axis swings slowly around instead, the way a leaning top circles while gravity tries to topple it. It traces a cone and comes back to its start in about 26,000 years [1]. The celestial equator is fixed to that axis, so as the axis swings, the equator swings with it. Its two crossings with the Sun’s path then slide along that path, through the constellations that lie on it. The first point moves west by about 50 seconds of arc a year [1]. A degree is a 360th of the circle and holds 3,600 seconds of arc, so the point covers one degree in about 72 years [2]. That rate roughly fits what Hipparchus saw. Two degrees take about 144 years, and Timocharis died about 130 years before Hipparchus defined the point, so his measurements were at least that old. A lifetime sees about one degree of the drift, which is too little to notice without instruments. But the drift never stops, and since Hipparchus it has carried the point roughly 30 degrees, a twelfth of the circle [2]. ## When the point left the Ram To say when the point left the Ram, the Ram needs an edge. A constellation used to be a pattern of stars with vague edges. In 1922 the International Astronomical Union fixed the list of constellations. By 1930 it had published exact borders for all 88, drawn by Eugène Delporte [4]. The borders run along lines of the sky’s longitude and latitude, so each constellation is now an area of sky with exact edges [4]. With exact edges, the date of the crossing can be computed. The astronomer Jean Meeus found that the point passed from Aries into Pisces in 68 BC, about sixty years after Hipparchus defined it [5]. So for all but those sixty of the roughly 2,150 years since, the first point of Aries has stood outside Aries. ## Why the name stayed The name survived because of what it had come to mean. The first point of Aries is wherever the Sun crosses the celestial equator heading north, and that definition mentions no star. A slide against the stars cannot make it false, because the stars are not in it. So astronomers had nothing to correct. The point they needed was still the March crossing, and the old name still picked it out. Geography kept an old sky name the same way. The Tropic of Cancer was named when the Sun stood in the constellation Cancer at the June solstice [6]. The solstice Sun has since moved on through Gemini and in 1990 entered Taurus [7]. Yet the line still means what it meant, the farthest north the Sun can stand directly overhead at noon [6]. In both cases the name was given for a view and kept for a definition. ## Is it still the sky’s zero? The name had nothing to correct, but the zero did. A zero that slides one degree in 72 years is awkward for precise work, because a position counted from it goes out of date. So astronomers replaced the zero and left the name where it was. Since January 1, 1998, star catalogs have used a frame fixed by quasars [8, 9]. Quasars lie far beyond the Milky Way, so far that their own motions across the sky should be negligible [9]. The frame’s zero of right ascension was set close to where the equinox stood in the year 2000 [9]. Measurements put it less than a tenth of a second of arc from there, which is less than the equinox slides in a day [9, 10]. The equinox itself keeps sliding west, so two things that began as one have come apart. One is the fixed zero of the catalogs, which the standard calls only the origin of right ascension [9]. The other is the moving point where the Sun crosses the equator each March, and that point keeps the name [2]. ## Why not name it for what it is? A better name would have no constellation in it. Name the point for its definition, and no drift can leave the name behind. The point already has such a name, the vernal point [2]. The vernal equinox is the spring equinox [3]. That rival is right about what a name should do. It says what the point is, and it will be as true in a thousand years as it is now. Its cost is a second name for one thing. The charts carry ♈ and the older books say Aries, so a reader of both must learn that the two are one. “Vernal” also carries a view of its own, because March brings spring only to the Northern Hemisphere [3]. So the old name stays because it is the one in use, and this note does not claim it is the better name. It misleads only a reader who takes it for a description. Naming the point for Pisces instead would repeat the first mistake, since the point leaves Pisces in 2597 [5]. ## Check it against the sky The answer rests on one claim a reader can test, that the March crossing now lies in Pisces and lay in Aries when it was named. Set a planetarium program to the March equinox of this year, and the Sun should stand in Pisces. Set it to 200 BC, and the Sun should stand in Aries. If the Sun stands in Aries on both dates, this note is wrong. A table gives the same test on paper. EarthSky lists where the Sun entered each constellation in 2018, in degrees along its path counted from the first point. The Sun entered Pisces at 351.57 degrees and Aries at 29.08 degrees [11]. The first point is at zero by definition, and zero lies between those two entries, inside Pisces. The same table allows a check of the quoted figures. From 351.57 around through 360 to 29.08 is about 37.5 degrees, the width of Pisces along the Sun’s path. At one degree every 72 years, the point needs about 2,700 years to cross it. Meeus’s two dates for entering and leaving Pisces are 2,664 years apart [5], within 1.4% of that figure. That agreement is weaker than it looks. Meeus worked from the same borders and a rate close to this one. So the check can catch a slip in the figures quoted here, but it cannot catch a wrong border. Part of the small gap is rounding, since at 71.6 years to a degree the crossing takes about 2,690 years [1]. The rate of precession is also not quite constant, and the borders are a convention, so other borders would give other dates. The drift itself is not in doubt, and only a different set of borders would move the dates. Why the name stayed is less sure. It is an inference from how the point is defined, and none of the sources below records anyone deciding to keep the name. A record of such a decision, giving another reason, would overturn it. ## Keep A name can outlive its scenery. ## Sources 1. Axial precession. *Wikipedia*: Hipparchus needed the Moon as a reference, at a lunar eclipse precisely 180 degrees from the Sun, and is thought to have measured Spica from it; Spica’s longitude about 2 degrees less than in the measurements of Timocharis (320–260 BC), and the same motion in other stars; the stronger pull of the Sun and Moon on the nearer half of the equatorial bulge, toward the ecliptic plane, and no precession on a perfect sphere; the equinoxes move westward, opposite to the Sun’s yearly motion; about 50 seconds of arc a year, one degree in about 71.6 years, a full circle in about 26,000 years; the obliquity of about 23.4 degrees. [wikipedia.org](https://en.wikipedia.org/wiki/Axial_precession) 2. First point of Aries. *Wikipedia*: the location of the March equinox, also known as the cusp of Aries or the vernal point; defined by Hipparchus in 130 BC at the western extreme of Aries; marked ♈; the point from which right ascension is calculated; one degree every 72 years, roughly 30 degrees since Hipparchus. [wikipedia.org](https://en.wikipedia.org/wiki/First_point_of_Aries) 3. March equinox. *Wikipedia*: known as the vernal or spring equinox in the Northern Hemisphere; the point where the Sun crosses the celestial equator northwards is called the First Point of Aries; it passed from Taurus into Aries in 1866 BC, into Pisces in 68 BC, and will pass into Aquarius in 2597 (Meeus). [wikipedia.org](https://en.wikipedia.org/wiki/March_equinox) 4. IAU designated constellations. *Wikipedia*: the list adopted in 1922; Delporte’s boundaries for each of the 88, adopted in 1928 and published in 1930, along arcs of right ascension and declination. [wikipedia.org](https://en.wikipedia.org/wiki/IAU_designated_constellations) 5. EarthSky. When will the Age of Aquarius begin? Jean Meeus’s dates with the IAU boundaries of 1930: Aries to Pisces in 68 BC, Pisces to Aquarius in 2597. [earthsky.org](https://earthsky.org/human-world/when-will-the-age-of-aquarius-begin/) 6. WorldAtlas. Tropic of Cancer: named when the Sun stood in Cancer at the June solstice; the northernmost latitude where the noon Sun can be overhead. [worldatlas.com](https://www.worldatlas.com/geography/tropic-of-cancer.html) 7. EarthSky. Solstice sun near the Taurus-Gemini border: the June solstice Sun in Gemini in 1989 and in Taurus in 1990. [earthsky.org](https://earthsky.org/tonight/sun-passes-into-the-constellation-gemini/) 8. International Celestial Reference System and its realizations. *Wikipedia*: the first frame adopted as of January 1, 1998, its axes fixed by extragalactic sources, mainly quasars. [wikipedia.org](https://en.wikipedia.org/wiki/International_Celestial_Reference_System_and_its_realizations) 9. IERS Technical Note 36, chapter 2, Conventional celestial reference system and frame: axes fixed with respect to the quasars, extragalactic objects whose proper motions should be negligibly small; the system replaced its predecessor on January 1, 1998; the origin of right ascension recommended close to the dynamical equinox at J2000.0; lunar laser ranging puts it 55.4 milliarcseconds from the inertial mean equinox of J2000.0. [PDF](https://impacts.obspm.fr/content/chapter2/tn36_c2.pdf) 10. IERS ICRS Center. Definition of the system: an earlier measurement put the x-axis 78 ± 10 milliarcseconds from the mean equinox of J2000.0. [obspm.fr](https://hpiers.obspm.fr/icrs-pc/misc/icrs.php) 11. EarthSky. Dates of the Sun’s entry into each constellation of the zodiac, 2018: Pisces at 351.57 degrees, Aries at 29.08 degrees. [earthsky.org](https://earthsky.org/space/dates-of-suns-entry-into-each-constellation-of-the-zodiac/) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/one-ell-and-eye/ Markdown: https://barani.org/writing/one-ell-and-eye/index.md Published: 2026-11-05 Modified: 2026-11-05 Form: Note Topic: Typography Series: Type, figures and tools, part 2 of 5 ======================================================================== # Why should 1, l and I look different in type for children? Three characters an adult rarely mistakes can stop a beginning reader. Igor Barani · 5 November 2026 · 8 min read Typography · Note · Type, figures and tools, part 2 of 5 · https://barani.org/writing/one-ell-and-eye/ In some typefaces a lowercase l, a capital I and the numeral 1 are nearly the same upright stroke. An adult rarely notices, because the word around a character tells her which one it must be. A child who is still learning which shape is which has only the shape to go by. So type for children should give each look-alike a mark of its own. The l gets a tail, the 1 a flag and a foot, and the I a bar at each end. The zero is drawn narrower than the O, or marked inside where it stands alone. ## Why adults rarely notice In “Illinois” the first three strokes can only be a capital I and two lowercase ells, since no other reading makes a word. Adults rarely stumble, because the word around a stroke usually settles which it is [1]. That is why a typeface can draw the three almost indistinguishably [2] and still serve its adult readers. > **Figure 1.** The look-alikes, drawn alike and then each with its own mark. In the first row nothing tells the I, the l and the 1 apart, or the O from the zero. In the second, each has its own mark, so no two share an outline. The I has bars at both ends, the l a tail, and the 1 a flag and a foot. The zero is narrower than the O. A schematic drawn for this note, not a typeface. > > What the figure shows: Two rows of five shapes. In the first row, a capital I, a lowercase l and the numeral 1 are the same plain stroke, and the letter O and the zero are the same oval. In the second row each has its own mark: crossbars at both ends of the I, a tail at the foot of the l, a flag at the top of the 1 and a foot at its base, a round O and a narrow zero. Take the word away and even an adult is left guessing. A unit symbol stands alone after its number, and the symbol for the liter was a lowercase l. In 1979 the General Conference on Weights and Measures adopted a capital L as a second symbol for it. The stated reason was to avoid the risk of confusion between the letter l and the number 1 [3]. So the confusion was serious enough for the metric system to admit a second symbol, as an exception. ## What a beginning reader has to go on A child tells b from d by which side the round part sits on, and E from F by the extra bar at the foot. Eleanor Gibson, Anne Pick and two colleagues tested children aged 4 to 8 on telling letter-like forms from altered copies. Errors fell with age. The authors traced this to how children identify letters: they attend to the features that tell letters apart and learn little of the rest [4]. That finding explains why the look-alikes are hard. A capital I and a lowercase l drawn as the same stroke share every feature, so the child has nothing to discover. Only the surroundings can separate them, and a beginning reader gets the least help from the surroundings, because she is still working letter by letter. In arithmetic she loses even that help. An adult knows that everything in a sum is a numeral, so a lone upright stroke there must be a one. A child who is still learning which shapes are numerals cannot lean on that knowledge. She also meets the three shapes with nothing around them at all, on number charts and in the alphabet itself. All of this is an inference from Gibson and Pick’s finding, and no study read for this note tested it. ## A mark for each If children learn letters by their differences, each of the three needs a difference the child can see. Type designers already work this way. They tell 1, l and I apart by the small finishing strokes at the ends of the upright, called serifs. What counts is whether a serif is present, where it sits and which way it points [5]. That is what the tail, the flag and the bars do, and the tests below say which of them matter most. In a test of type for European road signs, about 150 people took part, reading characters aloud from a screen at a distance [6]. The characters 1, f, i, I, l and j were mixed up especially often with one another [6]. From the errors the researchers drew advice for each character. A bent end at the foot of the l was “of great importance” for telling it apart. A distinctive arm at the top of the 1 helped, and for the I they advised serifs at top and base [1, 6]. The foot of the 1 is where the evidence divides. The road-sign test, with adults, found that serifs at the bottom of its upright “do not seem to have influence” [6]. Thomas Bohm, reviewing the research on misread characters, reports a study by Barbara Chaparro and colleagues that found a strong effect. They showed people single characters from 20 onscreen typefaces, each for a brief moment [7]. By Bohm’s account, a 1 with a top arm and a base was named correctly 97% of the time, against 43% with the arm alone [1]. Only the abstract of that study could be read for this note, so the two figures rest on Bohm. So one test found no clear effect of a foot, and the other, by Bohm’s account, found a large one. A foot costs nothing to draw. The 1 should therefore have its flag and a foot as well. Andika, a typeface made for literacy work and beginning readers, is drawn on this principle. Its capital I, lowercase l and numeral 1 are designed not to look alike [8]. It also gives small distinguishing features to letters that mirror each other, which can also confuse a new reader [8]. ## The zero and the O The zero and the capital O are the other common pair, and the usual difference between them is width. In most typefaces whose characters vary in width, the zero is a narrow ellipse and the O a broader, nearly round form [5]. Width is a weaker mark than a tail or a flag, because it shows best by comparison. A zero standing alone has no O beside it. A child can learn narrowness as a feature, but she has to carry the other shape in her head to use it. A mark inside the zero works without the comparison. Faces that give every character the same width cannot use width at all. Their designers add a slash or a dot instead, almost always to the zero [5]. So the choice is a trade. The narrow zero adds nothing to learn and helps most where the two shapes can be compared. The slashed or dotted zero can be read alone and costs one more mark to learn. This note keeps the narrow zero for ordinary reading. It would choose the marked zero wherever a zero must be read by itself, as on a number chart or in a code. Which serves beginners better was not tested in anything read for this note. ## Do special letterforms help at all? Ordinary type with good teaching may serve a child as well, and that view has evidence behind it. Books for young readers often use the simpler a and g of handwriting, called single-story or infant forms [9]. Those are the shapes children learn to write [10]. Researchers at the University of Reading tested typefaces with and without those forms on 48 children. They found no statistically significant differences in reading between the faces [9]. A test of 48 children can miss a small effect, but it would probably have caught a large one. Some children called one form of a and g harder than the other, and it did not affect how they read [9]. Ann Bessemans, working with children aged five to ten, found results like Walker’s. The children did not seem to have problems with ordinary letterforms. Reviewing earlier studies, she wrote that there seems to be good evidence the special characters are not always as legible as claimed [11]. If that view is right, a publisher gains a wider choice of typefaces. A child also reads from the start the letters she will read all her life. That evidence weighs against the special a and g, but it weighs less against separating the look-alikes, for two reasons. First, an a looks like no other letter in either form, so the child always has a feature to learn. An l drawn like a 1 gives her none. Second, the Reading test measured children reading words aloud [9], and a word carries the context that settles which stroke is which. A lone 1 on a number chart or an l after a number carries no such context. There, the only difference between two characters is the one the typeface draws. ## Check it across the room Print I, l, 1, O and 0 each on its own card, in the typeface and at the size the child will read. Shuffle the cards, hold one up across the room, and ask someone to name it. Any card named wrong, or named after a pause, needs another mark. A test that could go against this note uses children and two real typefaces. One is an ordinary face for adults in which the I and the l differ only slightly. The other gives each character its own mark. Show beginning readers single characters from each, one at a time and with no word around them, and count the errors and the hesitations. If the counts come out the same, the marks matter less to children than this note claims. What neither test can show is how much the confusion costs a child in real reading. There, size, weight, spacing and lighting also count, and so does her age. The evidence about which marks help comes from adults reading signs and screens. That adults confuse these characters is measured. That beginning readers are hurt more is the inference marked above. The second test above would settle it, and a finding of no difference would mean the title’s “should” is too strong. ## Keep Give each look-alike a mark of its own. ## Sources 1. Bohm T. Letter and symbol misrecognition in highly legible typefaces for general, children, dyslexic, visually impaired and ageing readers. *Typography.Guru*, 2019: the findings of Smuc and colleagues matched to the l, the 1 and the I, and the result of Chaparro and colleagues (2011) on the base of the 1, 97% against 43%. [typography.guru](https://typography.guru/journal/letters-symbols-misrecognition/) 2. Hackaday. Ask Hackaday: what goes into a legible font, and why does it matter? December 22, 2025: the capital I, lowercase l and numeral 1 can be almost indistinguishable in some fonts. [hackaday.com](https://hackaday.com/2025/12/22/just-what-goes-into-a-legible-font-and-why-does-it-matter) 3. Bureau International des Poids et Mesures. 16th General Conference on Weights and Measures, 1979, Resolution 6, Symbols for the litre: L adopted to avoid the risk of confusion between the letter l and the number 1. [bipm.org](https://www.bipm.org/en/committees/cg/cgpm/16-1979/resolution-6) 4. Gibson EJ, Gibson JJ, Pick AD, Osser H. A developmental study of the discrimination of letter-like forms. *Journal of Comparative and Physiological Psychology* 55(6):897–906, 1962: children aged 4 to 8; errors fell with age; little learning of the features that are not critical for telling letters apart. [doi:10.1037/h0043190](https://doi.org/10.1037/h0043190) 5. Bigelow C. Oh, oh, zero! *TUGboat* 34(2), 2013: the narrow zero and the round O; the zero receives the added slash or dot; one, ell and capital I told apart by the presence, location and orientation of serifs. [PDF](https://tug.org/TUGboat/tb34-2/tb107bigelow-zero.pdf) 6. Smuc M, Windhager F, Siebenhandl K, Egger S. IN-SAFETY Impaired Visibility Typeface Test report, 2007: 150 participants (those analyzed aged 19 to 66) read characters from a screen at three distances; 1, f, i, I, l and j were mixed up; a bent terminal on the l, a distinctive arm on the 1, serifs on the I, and no clear influence of serifs at the foot of the 1. [PDF](http://luc.devroye.org/IIID-Tern-Study-2009.pdf) 7. Chaparro BS, Merkle EC, Fox DE, Chaparro A. Examination of the legibility of isolated characters of onscreen typefaces. *Information Design Journal* 19(1):28–45, 2011. Abstract only, from the Wichita State University repository: characters from 20 onscreen typefaces identified after a brief exposure. doi:10.1075/idj.19.1.04cha. [wichita.edu](https://soar.wichita.edu/items/ab2dc3df-edb2-4e6e-96f5-c5b59c6a25e1) 8. SIL International. Andika: design. A typeface for literacy and beginning readers, with a capital I, lowercase l and numeral 1 that do not look alike, and marks that separate mirror-image letters. [sil.org](https://software.sil.org/andika/design/) 9. Walker S. Typography in children’s books. *Books for Keeps* 154, September 2005: the University of Reading tests with 48 children reading aloud; no significant differences between typefaces, with or without infant a and g. [booksforkeeps.co.uk](https://booksforkeeps.co.uk/article/typography-in-childrens-books/) 10. Strizver I. Typography for children. *Fonts.com (MyFonts)*: one-story a and g, the lowercase shapes children learn to write. [myfonts.com](https://www.myfonts.com/pages/fontscom-learning-fyti-situational-typography-typography-for-children) 11. Bessemans A. Typefaces for children’s reading. *TMG Journal for Media History* 19(2), 2016: children aged five to ten did not seem to have problems with non-infant characters; from earlier studies, good evidence that the special characters are not always as legible as asserted. [doi:10.18146/2213-7653.2016.268](https://doi.org/10.18146/2213-7653.2016.268) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/serif-and-sans/ Markdown: https://barani.org/writing/serif-and-sans/index.md Published: 2026-11-05 Modified: 2026-11-12 Form: Essay Topic: Typography Series: Type, figures and tools, part 3 of 5 ======================================================================== # Is matching x-heights enough to pair a serif and a sans? Two typefaces and two voices, at one size. Igor Barani · 5 November 2026 · revised 12 November 2026 · 20 min read Typography · Essay · Type, figures and tools, part 3 of 5 · https://barani.org/writing/serif-and-sans/ My children’s books speak in two voices, even within one line. The book’s own text is set in a serif, a typeface with small finishing strokes at the ends of its main strokes. The child’s work is set in a sans, which has none. The cheapest way to pair them is to shrink the sans until its x stands as tall as the serif’s x. That is not enough, because the sans’s H and its 7 then fall short. No change of size can close that gap, because a change of size lifts capitals and lowercase together. With EB Garamond and Hanken Grotesk, the two faces my fonts start from, the sans’s capitals would stand about 12 percent short and its digits 7.9 to 10.8 percent. Only a new drawing changes a face’s proportions, so the serif’s lowercase was redrawn taller and the sans was fitted to it. Both faces were changed under my direction. In the faces that resulted, Barani Serif and Barani Sans, the lowercase heights agree within 2.3 percent and the upright strokes within 3 percent. The price is capitals 4.4 percent shorter in the sans, and digits 2.2 to 3.6 percent shorter. That readers will feel the pair is better matched is reasoned, not yet tested. ## A change of size keeps a face’s proportions The gap can be measured, because every letter is drawn on a square. A type size names the height of that square, called the *em*. Both faces divide the em into 1,000 units, counted up from the *baseline*, the line the letters stand on. The height of the lowercase is called the *x-height* and is measured at the flat top of the x [1]. Set at one size, as they come, Hanken Grotesk’s x-height is 493 units and EB Garamond’s is 405, so the sans’s lowercase is 22 percent taller. The capitals disagree less. They rise to the *cap height*, measured at the top of the H, and the sans’s is 7 percent greater. Each upright stroke of the n is a *stem*, and in Regular Hanken’s stems are 85 units thick against EB Garamond’s 70, so they are 21 percent thicker. The two faces disagree in their proportions because they were drawn by different hands for different work. EB Garamond is Georg Duffner’s revival of Claude Garamond’s roman, from a specimen printed in 1592, later extended by Octavio Pardo [2]. Hanken Grotesk, a project led by Alfredo Marco Pradil, is a sans “inspired by the classic grotesques,” meant for work from signage to books, interfaces and websites [3]. Neither was drawn to sit beside the other, so nothing holds their proportions together. The measurements suggest a match that needs no new font. Shrink Hanken Grotesk to 405/493 of its size, about 82 percent, and in Regular its x-height is 405 units and its stems about 70, both EB Garamond’s own. The same result comes from setting the unchanged sans at 8.2 points beside the serif at 10. The capitals refuse. EB Garamond’s cap height is 1.61 times its x-height and Hanken’s only 1.41. A scale multiplies every height in a face by one number, so it leaves that ratio as it was. So a scale that matches the lowercase leaves the sans’s capitals at about 573 units against the serif’s 653, about 12 percent short. Hanken’s digits stand as tall as its capitals, so the digits fall short as well. They fall short by less, because the serif’s own digits stand below its capitals. EB Garamond’s 7 is 626 units tall against its H at 653. The scaled sans’s 7, at about 573 units, is therefore 8.5 percent short of the serif’s 7. Across the ten digits the shortfall runs from 7.9 to 10.8 percent. These numbers were computed for this essay from the heights and stems in the font files. Figure 1 plots the three heights beside those of the finished faces. > **Figure 1.** Three heights in units above the baseline, the x-height, the top of the 7 and the cap height, each line running from a serif (dot) to a sans (ring). Left, EB Garamond beside Hanken Grotesk scaled to 82 percent, so the x-heights match: the 7 stands 8.5 percent short and the H 12.3 percent. Right, Barani Serif beside Barani Sans at one size: the x-height 2.3 percent taller, the 7 2.2 percent short and the H 4.4 percent. One scale serves both panels and starts at 380 units. Heights measured from the font files. > > What the figure shows: Two slope charts of three heights, in units above the baseline, each line running from a serif to a sans. Left, EB Garamond beside Hanken Grotesk scaled to 82.2 percent so the x-heights match: x-height 405 and 405, 0%; the 7 626 and 573, −8.5%; the H 653 and 573, −12.3%. Right, Barani Serif beside Barani Sans at one size: x-height 440 and 450, +2.3%; the 7 650 and 636, −2.2%; the H 665 and 636, −4.4%. In both pairs the sans's 7 stands as tall as its H. The ratio changes only when the lowercase and the capitals move by different proportions. No scale can do that, and a new drawing can. In Barani Serif the lowercase stands at 440 units and the capitals at 665, so the serif’s cap height is 1.51 times its x-height, closer to Hanken’s 1.41. At equal x-heights, that change alone would cut the capitals’ gap from about 12 percent to about 6.5. ## The voices should differ in shape and agree in size The books are With Every Reason, a series in arithmetic for my children and for all children. The serif sets the reasons, which are the text a child reads, the Guides for parents and teachers and the worked solutions. The sans sets the work, which is the problems and exercises, the tables and the figure labels, at the serif’s size. By the books’ typography standard, two voices let the eye tell reading from doing before it reads a word, a claim no test with children has checked. My font specification records my words that I really like Garamond and that the books need a better sans, chosen for its strength in mathematics and formulas. I asked for faces that look as if one designer drew them, matched in x-height and in the other measures of size and spacing wherever they can be. Barani Serif and Barani Sans are built from two open-source typefaces, EB Garamond and Hanken Grotesk. The height matters, because the eye seems to judge a face’s size by its x-height. Karow measured 1,049 roman typefaces and found x-heights from 0.28 to 0.58 of the em, as Legge and Bigelow report in their review [4]. EB Garamond’s is 0.405 and Hanken’s 0.493. Faces with larger x-heights “appear to be bigger,” they write, although this is “not proven by rigorous laboratory studies” [4]. If it holds, the child’s work set at 10 points looks as large as the text beside it would at 12.2 points. Taller letters with thicker stems put more ink on the line. Squint at a page and its letters blur into a gray, which typographers call the page’s *color* [5]. This essay measures color as the share of a line’s area that ink covers, with the 26 lowercase letters set in a line one em tall. By that measure Hanken’s line is 27 percent darker than EB Garamond’s. Allan Haley advises pairing faces that differ. “The more dissimilar the type designs, the better the mix,” he writes, and he counts a serif with a sans as the least risky way to mix faces from different families [6]. My books want that difference in shape, serifs against none. They must avoid a difference in size or color, because type uses both to mark emphasis. Where the voices meet in one line, a jump in either would, by this reasoning, make the child’s work look louder than the text around it, a signal nobody meant. The heights, the stems and the color are marked in Figure 2 on the finished Barani faces, where the measures nearly agree. In Regular, the stems of the n are 70 units in the serif and 69 in the sans. Ink covers 17.1 percent of the line in the serif and 16.9 percent in the sans. > **Figure 2.** The parts of a letter, in Barani Serif and in Barani Sans at one size. Above, the cap height, x-height and baseline run across H, n, d, p and o. In the middle, the dotted boxes of n and o meet edge to edge, while kerning, a correction for one pair, makes the boxes of T and o overlap. Below, a bar one em tall holds the ink of a line from a to z spread evenly, so its gray is the line’s color. Outlines from the font files. > > What the figure shows: Barani Serif Regular: the letters H, n, d, p and o set apart, with three lines across them: the cap height at the top of the H, 665 units above the baseline; the x-height at the top of the lowercase, 440 units; and the baseline the letters stand on. Labels mark the serif at the foot of the H, a stem of the n and the counter between its stems, the ascender of the d, which rises to 715 units, and the descender of the p, which falls 285 units below the baseline. The o is thick at its sides, 82 units, and thin at its top, 31 units. Below, the pair n o as its spacing sets it: the two letters' boxes, dotted, meet edge to edge, and the white between the inks comes from both. Then T o, where kerning moves the o 105 units left, so the boxes overlap. Last, a to z over a bar one em tall holding the same ink spread evenly: ink covers 17.1 percent of it, the line's color. Barani Sans Regular, at the same size and in the same layout: the cap height 636 units, the x-height 450. The H has no serifs, and the o's stroke is nearly even, 68 units at the sides and 60 at the top. The pair n o meets edge to edge; in T o, kerning moves the o 88 units left. Ink covers 16.9 percent of the bar under a to z. The faces are matched on the n’s stems and not on the o, because the serif’s o is meant to swell. It is 82 units thick at its sides and 31 at its top, and that difference between thick and thin is the *stroke contrast* [5]. The sans’s o measures 68 units at its sides and 60 at its top, so its stroke is nearly even. In these books the capitals’ gap falls on the numbers. The sans sets the sums, and its digits stand at its cap height, in Hanken Grotesk and in Barani Sans alike. So wherever the child’s work holds a number, a shortfall in the sans’s capitals carries into that number, though by less. ## The serif’s lowercase is raised, not scaled On October 26 I confirmed a plan to lift the serif’s x-height from 405 to 440 units by redrawing its lowercase, not by scaling it. Scaling EB Garamond up by 8.6 percent would also reach 440, but it would thicken every stroke, widen every letter and lift the cap height from 653 to about 709. The ratio of capitals to lowercase would stay at 1.61. So the lowercase was redrawn instead, and the n shows how. > **Figure 3.** Barani Serif’s n, each point of its outline marked against the same point of EB Garamond’s n, which is not drawn. Rings rose and dots stayed, so the top of the letter moved as one piece and only the stems grew longer. The lines mark the two x-heights. Points measured from the font files. > > What the figure shows: The lowercase n of Barani Serif, with the 92 points that fix its outline, its curve handles included. Compared with the same points of EB Garamond's n, which is not drawn, 41 points rose, 40 of them by 35 units and one by 36: the arch, its joins and the serif at the top. The other 51 stayed where they were: the two feet with their serifs and the lower part of each stem. A solid line marks Barani Serif's x-height at 440 units and a dotted line EB Garamond's at 405. The outline of the n is fixed by 92 points, the handles that steer its curves included. Of these, 51 stay where they were, 40 rise together by 35 units and one curve handle rises 36. Because the top moves as one piece, the arch, its joins and the serif at the top keep their shapes to within a unit, and only the stems grow longer. The m behaves the same way, in both the Regular and the Bold weights. So the x-height goes from 405 to 440 units in Regular and from 425 to 460 in Bold, while the stems keep their thickness. The capitals rose too, by 10 to 12 units, so the cap height is 665 in every weight. An *ascender* is a stem that climbs past the x-height, as in b, d and l. The ascenders rose 10 or 11 units, to 715, so the capitals still stand a little below them, as in Garamond. ## The sans is fitted to the serif, and its capitals pay On October 27 I reviewed a proof of the serif’s new x-height and judged that serif and sans did not match. The sans was then fitted to the raised serif. A scale for each weight was tried first, to meet the serif’s x-height at 440 units in Regular and 460 in Bold. But the Bold’s larger scale made its capitals and digits 4.5 percent taller than the Regular’s, a step in height at every change of weight, so one scale now serves every weight. It shrinks Hanken Grotesk to 450/493 of its size, or 91.3 percent. The sans’s x-height is then 450 units, halfway between the serif’s two. It stands 10 units above the serif’s Regular, which is 2.3 percent, and 10 units below its Bold. The stems came next. Hanken Grotesk’s weight can be set anywhere along a continuous scale, not only at its named weights [3]. The sans’s Regular was chosen by eye from three trial settings, which fixed the ratio of its stem to the serif’s at 69 units to 70. At the final scale, every weight, the Regular included, was set to keep that ratio. Measured across both stems of the n, the sans’s stems come out 0.8 to 2.7 percent thinner than the serif’s, weight for weight. Matched stems bring the color close in Regular but not in Bold. Measured over a line of a to z, as in Figure 2, the Regular sans comes within 1 percent of the serif’s color, but the Bold sans is 8 percent darker. No measurement made for this essay explains that difference. The capitals still pay, though less than under the cheapest match. The raise brought the two ratios closer, at 1.51 and 1.41, and did not make them equal. So one scale can still match only one of the two heights. The finished sans follows the x-height, as the cheapest match does. At equal x-heights the capitals’ gap would be the 6.5 percent already worked out. The sans’s lowercase stands 2.3 percent above the serif’s Regular and lifts its capitals with it. So the sans’s capitals stand 636 units tall against the serif’s 665, which is 4.4 percent shorter. Its 7 stands 636 units tall against the serif’s 650, which is 2.2 percent shorter. Across the ten digits the shortfall runs from 2.2 to 3.6 percent. ## One construction draws the signs Building new fonts brings a gain that no scale can. A scale leaves the math signs, such as + and =, and the white between characters as each face’s designer left them. Once the faces are being changed, one written set of rules can govern both. Even inside one face, signs can disagree. In EB Garamond, the + centers 16.5 units lower than the =, so the signs of a sum step up and down along the line. In the Barani faces, one construction program draws + − × ÷ and =, each face’s signs with that face’s own pen. In the serif, every sign measured for this essay centers 252 units above the baseline. That height is the *math axis*, the line on which formulas center their signs and fraction bars. > **Figure 4.** Hyphen, minus, plus and equals in Barani Serif and in Barani Sans, each row on its own axis. The serif’s signs end in the slanted, rounded tips of a broad pen, and the sans’s in square ends with an even stroke. The sans keeps a lower axis, so the top of its minus is level with the bottom of its hyphen. Outlines from the font files. > > What the figure shows: Hyphen, minus, plus and equals in Barani Serif and in Barani Sans. The serif's signs end in the slanted, rounded tips of a broad pen; the sans's end square, with an even stroke. The serif's signs center on an axis 252 units above the baseline. The sans's center on 226, so its minus ends at the height where its hyphen begins. The sans keeps a lower axis, to hold its minus clear of its hyphen. Its hyphen, Hanken’s own scaled with the face, sits high, from 253 to 304 units above the baseline. A minus centered on the serif’s axis would span 225 to 279 units, by this essay’s arithmetic, and so overlap the hyphen by about half its thickness. So the sans’s minus stays on Hanken’s own axis, 226 units, where it spans 199 to 253 units and its top just meets the hyphen’s bottom. The two bars then share no height, which should keep a child from taking one for the other. The cost is a step at the change of voice, since the sans’s signs sit 26 units lower than the serif’s. The same rules keep look-alike characters apart in both faces, with a zero about six-tenths as wide as the O and an open 4. The note on [why 1, l and I should look different](https://barani.org/writing/one-ell-and-eye/) gives the reasons. ## A floor under the white The same rules put a floor under the white between neighbors in each face. The crowded pairs this section shows come from the project’s own builds, so the floor is a check on the new fonts, not a repair of the stock ones. Each letter is set in a box that holds its ink and, in most letters, some white on both sides, so that letters set side by side stand evenly apart. That built-in white is the *spacing*, shown between n and o in the middle of Figure 2. A few pairs still sit badly, such as T then o, where spacing alone would leave the o far out from the T’s stem. The font then carries a correction for that pair alone, called *kerning* [1]. In Figure 2, kerning moves the o 105 units left, under the T’s arm, in the serif, and 88 units in the sans. Kerning can also push a pair apart, and one pair showed the need. In the serif’s first release candidate, a build put forward to become version 1.0, the equals sign touched a following gamma. A gate, a program that checks every build and stops it when a check fails, now catches such pairs. It measures the shortest distance between the inks of every pair of neighbors it checks, corners included. No pair may come closer than the floor, 25 units in the serif and 26 in the sans. A pair under the floor gets kerning, and this kerning only ever adds white. > **Figure 5.** Equals followed by gamma in Barani Serif Regular, enlarged where the two meet, with a gray band marking the floor 25 units out from the ink of the equals sign. In the first release candidate the gamma’s arm crosses the band and touches the sign. In the third, kerning moves the gamma 45 units right, and its arm stops at the band’s edge, 25.2 units from the sign. Measured from the outlines as the font spaces them. > > What the figure shows: Equals followed by gamma in two release candidates of Barani Serif Regular, each shown whole and then enlarged where the two meet, with a gray band 25 units wide around the equals sign, the floor. In the first candidate the gamma's arm crosses the band and touches the equals sign. In the third, kerning moves the gamma 45 units to the right, and the shortest white between them is 25.2 units, so the arm stops at the band's outer edge. The floor comes from a model of the press, not from a printed sheet. The model takes the worst case among the presses the books might use, 1,200 dots per inch, with toner spreading 21 micrometers on each side of a stroke. It sets type at 9.5 points, the size of the Guides’ notes. A program moved the digits of each fraction, such as ½, away from the slash until nothing joined, and the largest gap needed, rounded up, became the floor for every pair. At 9.5 points, the serif’s 25 units come to about 84 micrometers. Both faces take their floor from that one model, so the white in a serif sentence and in a sans sum should survive printing alike. The same gate opened other crowded pairs, such as 4 followed by + in Bold, from 3.0 to 27.0 units of white. The floor sets a minimum and leaves the rest of the spacing alone, because extra-wide spacing did not speed reading in a study that Legge and Bigelow cite [4]. ## How the fonts were made A redrawing can go wrong where a scale cannot, so the case for redrawing depends on how the drawings were checked. I directed the work. Letters and signs were drawn by rule, and tools were built to check them. The drawings were judged by comparing alternatives at 200 points and at reading size, often blind, each labeled only by a letter. Every new gate was first shown to fail on an older build, as the floor’s gate failed on the serif’s first release candidate. One measuring program failed that discipline. Asked which side of a slashed sign to measure, it looked under the wrong name, found no entry and silently measured the left side. A blind choice built on its numbers was withdrawn. Now the programs share one function for each measure, and a lookup that finds no entry stops with an error. ## The case for scaling the sans The strongest case against this work is that it may not be needed. The cheapest match, the unchanged sans set at 82 percent of the serif’s size, fits the lowercase and the Regular stems with no font changed. That rival has real gains. No font is changed, so EB Garamond stays as it was revived from the specimen of 1592 [2]. A size is a setting, so the match can be undone at any time. Its Regular lowercase heights agree exactly, where the finished faces differ by 2.3 percent. Its best argument concerns reading. In this essay’s own text, 97 percent of the letters are lowercase, about 16,900 against fewer than 500 capitals, so the scale follows the x-height. The finished faces rest on the same argument, since their one scale follows the x-height too. Under that rival, the sans’s capitals stand about 12 percent short where the finished faces cost 4.4. The rival sets two sizes at every change of voice. It also leaves each face’s signs and white as they come, with EB Garamond’s + still 16.5 units below its =. The answer still stands because of where the cost lands. The argument from the lowercase holds for the serif, which sets sentences. The sans sets the problems, the exercises and the tables, where a line may hold more digits than lowercase letters, and the digits stand at the cap height. So the rival sets the digits of a sum 7.9 to 10.8 percent shorter than the digits in the text beside them. The finished faces set them 2.2 to 3.6 percent shorter. A scale only moves the gap, onto the digits when it matches the lowercase and onto the lowercase when it matches the digits. A redrawing shrinks the gap itself, and the raise cut the capitals’ gap at equal x-heights from about 12 percent to 6.5. The sans sets words as well as sums, so the gap had to shrink, not move, and that is why I chose a redrawing over a scale. A second case is that a family drawn as one can agree from the start. Bigelow & Holmes’s Lucida faces, serif and sans among them, all share “the same x-height, capital height, and similar series of weights” [7]. Equal heights give equal ratios, so in such a family one size fits capitals and lowercase alike. Barani Serif and Barani Sans began as two faces by different designers, so their agreement had to be made afterward. The cost of that rival is the serif itself, since a book set in Lucida is no longer set in Garamond. A third case is the advice to pair faces that differ. Haley warns that “sans serifs that are similar in design almost never work well together” [6]. His warning concerns two sans faces, so applying it to a serif and a sans fitted to each other goes beyond what he wrote. But the risk it points to is real, since such a pair could read as one face that is slightly off. A reply may be that two things still separate the voices, the serifs and the stroke contrast. If they do, they alone mark the child’s work. It is a judgment, and no test with readers has checked it. ## Check The measured half of the answer can be tested by anyone who has two font files. Divide each face’s cap height by its x-height. If the two ratios are equal, one size fits both heights, and matching x-heights is enough for that pair’s heights. The test says nothing about stems, color, signs or white. Lucida’s faces are such a pair [7]. For EB Garamond and Hanken Grotesk the ratios are 1.61 and 1.41, so the test could have failed and did not. Those ratios are the scaled heights in another form, since 1.41 divided by 1.61 and 573 divided by 653 are one calculation, and it gives 0.88, the 12 percent gap. The costs above are measured, but the benefit still waits on blind tests with parents, teachers and typographers, and on a press sheet on the books’ paper. Six things count against the work. The Bold sans is 8 percent darker than the Bold serif although its stems match. The lowercase heights now differ by 10 units, 2.3 percent in Regular, where the cheapest match fits the Regular exactly. The sans’s signs sit 26 units lower than the serif’s, a step at every change of voice. The premise that the eye judges a face’s size by its x-height is, in Legge and Bigelow’s words, “not proven by rigorous laboratory studies” [4]. The books’ claim that two voices help a child tell reading from doing has not been tested with children. And how the sans’s work divides between words and sums was not counted, so how much the rival’s shortfall in the digits would be seen is reasoned, not measured. Matching is not new, as Lucida shows [7]. No survey of other matched families was run, so I make no claim to be first. The essay leaves out the italics, which are still being built, the small capitals, Barani Math, a third face that sets formulas, and the cost of building the fonts. The Barani fonts are not yet released to anyone else. The books and this site are set in them, and how they will be offered to others is still to be decided. This essay was drafted from the project’s logs and from the font files. A script measured again every number taken from the font files, and every outside fact was checked against a source retrieved for the essay. The numbers measured from the font files can be held firmly, since a script reproduces each one. That a reader feels the difference can be held only as a reasoned expectation, since no reader has been tested. For the serif, my font specification already names the result that would change my mind. Its seventh principle keeps a departure from EB Garamond “only if it beats or equals it in blind comparisons printed on the books’ paper”. The same test is proposed here for the sans and for the matching. So if readers comparing blind on that paper prefer pages set in EB Garamond and Hanken Grotesk, the matching is not kept. If they cannot tell the two apart, it stays. ## Keep A scale only moves the gap; a redrawing shrinks it. ## Sources 1. Phinney 2025, Phinney on Fonts. Typography glossary: baseline, x-height, cap height, contrast, kerning, the em, and advance width with its sidebearings. [thomasphinney.com](https://www.thomasphinney.com/typography-glossary-for-document-examiners-others/) 2. Wikipedia. EB Garamond, a revival by Georg Mayr-Duffner (credited in the fonts as Georg Duffner) of Claude Garamond’s roman from the Berner specimen of 1592, continued by Octavio Pardo. [wikipedia.org](https://en.wikipedia.org/wiki/EB_Garamond) 3. The Hanken Grotesk Project Authors. Hanken Grotesk, README, led by Alfredo Marco Pradil. [github.com](https://github.com/marcologous/hanken-grotesk) 4. Legge & Bigelow 2011, *Journal of Vision* 11(5):8. Does print size matter for reading? A review of findings from vision science and typography. Reports Karow 1993 on 1,049 roman typefaces, and Yu et al. 2007 on extra-wide spacing. [PubMed Central](https://pmc.ncbi.nlm.nih.gov/articles/PMC3428264/) 5. Google Fonts Knowledge. Glossary entries for color, counter, contrast and spacing. [fonts.google.com](https://fonts.google.com/knowledge/glossary) 6. Haley, Fonts.com (Monotype). Mixing typefaces. [myfonts.com](https://www.myfonts.com/pages/fontscom-learning-fontology-level-2-making-type-choices-mixing-typefaces) 7. Bigelow 2005, TeX Users Group. Notes on Lucida designs. [tug.org](https://www.tug.org/store/lucida/designnotes.html) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/figures-for-the-projector/ Markdown: https://barani.org/writing/figures-for-the-projector/index.md Published: 2026-11-05 Modified: 2026-11-05 Form: Note Topic: Design Series: Type, figures and tools, part 4 of 5 ======================================================================== # How should a figure be drawn for the projector? A chart that looks sharp at the desk can fade in a lit hall. Igor Barani · 5 November 2026 · 9 min read Design · Note · Type, figures and tools, part 4 of 5 · https://barani.org/writing/figures-for-the-projector/ A gray line that is easy to read on a laptop can fade away on a projector. The hall’s lamps light the screen too, and they turn its black to gray. So draw the figure for that hall, and for eyes that miss color. Use near-black ink and lines at least 2 points thick on the slide. Choose colors that differ in lightness as well as hue, and write a name on every line. ## Why the hall steals contrast A laptop screen makes its own light, and where it shows black it gives off almost none. A projector can only add light to a screen and cannot take any away. So the darkest part of a projected image is the screen itself, lit by whatever lamps and windows the room has. The room sets the best black the projector can reach [1, 2]. The loss can be put in numbers. The usual measure of contrast compares how much light two colors give off, on a scale where black is 0 and white is 1. It adds 0.05 to each as an allowance for glare and divides the lighter by the darker [3]. Black on white therefore comes to 1.05 divided by 0.05, which is 21 to 1, the largest ratio the measure allows [3]. Now suppose the room’s light adds a tenth of white, 0.1, to every part of the screen, and call that the example hall. White rises from 1 to 1.1, which changes little, but black rises from 0 to 0.1. The ratio becomes 1.15 divided by 0.15, and black on white falls from 21 to 1 to under 8 to 1. Every other color on the slide is squeezed between that lifted black and that white, so every contrast shrinks with it. > **Figure 1.** The same slide on a laptop and in a lit hall. With room light at 10% of white, the black 2-point line falls from 21 to 1 to 7.7 to 1. The gray 1-point line falls from 4.5 to 1, just enough for text, to 3.5 to 1. As text it would now fail, though as a line it still clears 3 to 1. Move the Room light slider, from 0% to 30%, to try another hall. Drawn for this note with the WCAG formula. > > What the figure shows: The same slide twice, with a black line and a gray line on white. On the laptop the black line has a contrast of 21 to 1 against white and the gray line 4.5 to 1. In a hall where room light adds a tenth of white to the screen, the black line falls to 7.7 to 1 and the gray line to 3.5 to 1, below the minimum for text. A tenth is an assumption, and a hall has a figure of its own that can be measured. Project a slide that is half black and half white, with the lights as they will be for the talk. Then, from a seat, measure the light coming off each half with a meter that reads a small spot. In the example hall the black half is about an eleventh as bright as the white half, since 0.1 is an eleventh of 1.1. A brighter black half means a worse hall. ## What the usual minimums assume The web’s accessibility guidelines, known as WCAG, set the usual minimums. Ordinary text needs a contrast of at least 4.5 to 1 against its background, and large text 3 to 1 [4]. The parts of a graphic needed to understand it, such as each line in a graph, also need 3 to 1 [5]. Those ratios are computed from the colors in the file. The allowance for glare is fixed at 0.05, a value based on typical viewing flare, the ambient light an ordinary screen reflects [3]. A web page cannot know the room it will be read in, so a fixed allowance is the best it can do. A presenter can know the room. In the example hall, a mid gray that passes at 4.54 to 1 on the laptop falls to about 3.5 to 1. Text that met the minimum at the desk misses it on the screen. The same sum says how much to spare. A gray has to reach about 7 to 1 on the laptop to keep 4.5 to 1 in that hall. So the first rule for the projector is to aim well above the minimum. Text and important lines belong in near-black on a white slide, or near-white on a dark one. Light grays and pale tints are the first things the hall erases. ## How thick a line should be A figure made for a journal arrives at the talk with thin lines. Nature, for one, asks for lines and strokes between 0.25 and 1 point at the final size [6]. That rule suits a page held in the hand. A seat in the back row is another matter. A point is 1/72 of an inch, and a widescreen PowerPoint slide is 7.5 inches tall [7], so the slide is 540 points tall. A 1-point line is then 1/540 of the screen’s height, however large the screen. Take, for example, a back row six screen-heights away. From that seat the line spans about one minute of arc, which is a sixtieth of a degree. At twelve screen-heights it spans half that, so every width has to double. One minute of arc is the width of the strokes in the letters on the 20/20 line of an eye chart [8]. So from the back row a 1-point line is as thin as the finest detail a normal eye is expected to resolve. Visual acuity also gets worse as the contrast of what is seen goes down [9]. The example hall has already taken nearly two thirds of the contrast. A line that thin can still be seen. What goes first is probably what the reader needs from it, which is its color, its dashes and the gap between it and its neighbor. Each of those is a detail at the limit of the eye. That step is an inference from the eye chart, and no test of projected lines was read for this note. A 2-point line spans about two minutes of arc from the same seat, twice the finest detail, and that margin is the reason for the rule. The width that counts is the width on the slide, after the figure has been placed and stretched to its final size. A journal figure enlarged to fill the slide thickens its lines in proportion. The same figure placed small beside a block of text keeps them hair-thin. So the rule is 2 points on a 540-point slide, or 1/270 of the slide’s height on any other. ## Color that survives Thick, dark lines still have to be told apart, and color is the usual way. It is also the least dependable. About one man in twelve has a color vision deficiency, and the most common kind makes red and green hard to tell apart [10]. Among women of Northern European ancestry the common red-green form affects as many as one in two hundred [11]. So a hall of a hundred, half of them men, holds up to about four such people. Each may see a red line and a green line as one color. The hall adds its own trouble. Room light dilutes every color with the same white, and so it pulls the hues toward one another. Lightness is the difference that survives both the hall and the color-blind eye. Two colors that are equally light have a contrast of 1 to 1 with each other, on the laptop and in the hall alike. Only their hue separates them, and hue is what the hall dilutes. A dark color and a middle one keep some contrast wherever they are shown. Bang Wong gives the same advice, that colors are easier to distinguish when they vary in lightness and saturation as well as hue [11]. A grayscale copy of the figure shows at once whether they do. On a white slide, that means pairing a near-black color with a middle tone. One such pair is a dark navy (#003366) and a vermilion. The vermilion comes from a palette Wong offers, first drawn up by Kei Ito and Masataka Okabe [12]. Readers with red-green color blindness can tell its colors apart [11]. Against white the navy reaches about 12.6 to 1 and the vermilion about 3.9 to 1. In the example hall they keep about 6.3 and 3.1, enough for a line. Against each other they have about 3.3 to 1 on the laptop and keep about 2 to 1 in the hall. So they still read as a dark line and a lighter one. A third and a fourth line cannot all differ that much in lightness, which is why the next rule matters more than this one. ## Names on the lines Color alone still leaves a gap, and the guidelines name it. They ask that color never be the only way information is given [13]. A legend in the corner, matched to the lines by color alone, breaks that rule quietly. It ties each color to a name and leaves the reader to carry the match across the slide. A name written at the end of each line needs no color vision and no trip to the corner. The reader who sees two lines as one hue can still tell which is which. Names on the lines also survive the hall, because black text on white keeps the most contrast of anything on the slide. ## Why not vary the lines instead? The legend could stay if the lines themselves differed, one solid, another dashed and a third dotted. Shape needs no color vision and survives a grayscale printer. This site’s own figures use the idea, drawing a second quantity dotted beside a solid first one. Patterns work for two lines and strain at three. Each dash is a short, thin line, so a dashed line probably fades sooner than a solid one on a washed-out screen. No measurement read for this note says so. With more lines, the patterns become a code the reader must learn from the legend, and the trip to the corner returns. Names on the lines do the same job without a code. The two methods also combine well, since a figure with two quantities can draw one solid and one dashed and still name both. ## Check it from the back row The best test is the room itself. Project the slides with the lights at the level they will have during the talk, and look from the back row. If every line and label holds there, the figure is ready. If a line is lost although the rules above were followed, the room is brighter than the example hall, and its own numbers win. The room can also decide against these rules. Put 1-point lines and colors that differ only in hue beside 2-point lines and colors that differ in lightness. If both read as well from the back row with the lights on, the rules are stricter than that hall needs. When the room is not to be had, turn each slide to grayscale and stand back six times the height of the laptop screen. Every line should keep its own gray, and every label should read without effort. Lines that turn the same gray may look alike to a reader with red-green deficiency. The cost of the rules is room. Thick lines and written names crowd a slide, so it holds fewer lines, and a dense figure has to be split across two. The detail is not lost, because the thin-lined version stays in the paper. The rules say nothing of the projector’s brightness, the screen’s surface or the angle of view, which all change the contrast. Nor do they cover how large the type must be, which is the question of this site’s Typography page. The direction of the answer is certain, because room light always lowers contrast. The size is not, since the tenth of white is an assumption and not a measurement of any hall. A dark room with a bright projector would allow thinner lines and paler colors. ## Keep Draw for the back row, lights on. ## Sources 1. Paullin T. The big mistake. *ProjectorCentral*: ambient light from any direction compromises a projected picture’s contrast; the room’s contrast sets the best black level. [projectorcentral.com](https://www.projectorcentral.com/The-Big-Mistake.htm) 2. Barco Residential. The importance of ANSI contrast in high performance residential applications, white paper: even the smallest light leaking from a window, a lamp or another room is added to the measured dark level. [barco.com](https://assets.barco.com/m/7f458cbdce2ad3e4/original/ANSI-Contrast-pdf.pdf) 3. W3C. Understanding WCAG 2.2, SC 1.4.6 Contrast (Enhanced): the ratio (L1 + 0.05) / (L2 + 0.05), the 0.05 based on typical viewing flare, and the maximum of 21 to 1. [w3.org](https://www.w3.org/WAI/WCAG22/Understanding/contrast-enhanced.html) 4. W3C. Understanding WCAG 2.2, SC 1.4.3 Contrast (Minimum): 4.5 to 1 for text, 3 to 1 for large text. [w3.org](https://www.w3.org/WAI/WCAG22/Understanding/contrast-minimum.html) 5. W3C. Understanding WCAG 2.2, SC 1.4.11 Non-text Contrast: 3 to 1 for the parts of graphics needed to understand them, such as each line in a graph. [w3.org](https://www.w3.org/WAI/WCAG22/Understanding/non-text-contrast.html) 6. Nature. Final submission, figure guidelines: line weights and strokes between 0.25 and 1 point at the final size. [nature.com](https://www.nature.com/nature/for-authors/final-submission) 7. Microsoft Support. Change the size of your slides: widescreen 13.333 by 7.5 inches. [microsoft.com](https://support.microsoft.com/en-us/powerpoint/change-the-size-of-your-powerpoint-slides) 8. Snellen chart. *Wikipedia*: on the 20/20 line, letters subtend five minutes of arc and their strokes and gaps one minute. [wikipedia.org](https://en.wikipedia.org/wiki/Snellen_chart) 9. Pluháček F, Siderov J, Macháčová I. Crowding can impact both low and high contrast visual acuity measurements. *Scientific Reports* 12:16338, 2022: acuity worsened as letter contrast decreased. [doi:10.1038/s41598-022-20479-y](https://doi.org/10.1038/s41598-022-20479-y) 10. National Eye Institute. Color blindness: about 1 in 12 men have color vision deficiency; the most common type makes red and green hard to tell apart. [nei.nih.gov](https://www.nei.nih.gov/eye-health-information/eye-conditions-and-diseases/color-blindness) 11. Wong B. Points of view: color blindness. *Nature Methods* 8:441, 2011: as many as 8% of men and 0.5% of women of Northern European ancestry have the common red-green form; colors are easier to distinguish when they vary in lightness and saturation as well as hue; a palette of eight colors, vermilion (213, 94, 0) among them, that readers with red-green color blindness can tell apart. doi:10.1038/nmeth.1618. Author Correction, *Nature Methods* 20:1266, 2023 (doi:10.1038/s41592-023-01974-0): adds a link to the 2002 presentation by Kei Ito and Masataka Okabe that the palette draws on. [nature.com](https://www.nature.com/articles/nmeth.1618) 12. Okabe M, Ito K. How to make figures and presentations that are friendly to color blind people, 2002: slide 15 lists vermilion (213, 94, 0) among colors easier to identify. [jfly.uni-koeln.de](https://jfly.uni-koeln.de/html/manuals/pdf/color_blind.pdf) 13. W3C. Understanding WCAG 2.2, SC 1.4.1 Use of Color: color is not the only visual means of conveying information. [w3.org](https://www.w3.org/WAI/WCAG22/Understanding/use-of-color.html) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/typed-fields/ Markdown: https://barani.org/writing/typed-fields/index.md Published: 2026-11-05 Modified: 2026-11-05 Form: Note Topic: Artificial intelligence Series: Type, figures and tools, part 5 of 5 ======================================================================== # In what form should an AI model hand its answer to code? A model says “not urgent.” The program reads “urgent.” Igor Barani · 5 November 2026 · 8 min read Artificial intelligence · Note · Type, figures and tools, part 5 of 5 · https://barani.org/writing/typed-fields/ A model reads a referral letter and reports, in a sentence, that the patient is not urgent. The clinic's scheduling program searches that sentence for the word *urgent*, finds it, and books the patient as urgent. The model should have answered in typed fields. Those are a choice from a fixed list, a probability for each yes-or-no question, and a score whose levels are written out in words. A program can check each field before it acts, whereas it can only guess at what a sentence meant. ## Start from the decisions the program must make The clinic in that case is an example made up for this note. It receives referral letters from other doctors and wants a model to help sort them. The model reads every letter first, and a nurse still reviews the hard ones. What the model says goes to the scheduling program, which is the code of the title. The program must then do something, such as book a visit, ask the referring office for more, or flag the letter for the nurse. A program can act only on answers it was built to expect, so the first step is to name the decisions it must make. Here there are three. The program must know how soon to see the patient, whether images came with the letter, and whether the letter is complete enough to book from. Each question is narrow and each leads to an action, which is what makes it a good question for a program. "Summarize this referral" is a poor one, because no action follows from a summary. > **Figure 1.** One letter, read two ways. A search for the word *urgent* books the letter as urgent, although the sentence says “not urgent.” Each typed field leads to one action. An example drawn for this note, not from any clinic. > > What the figure shows: Above, the sentence “Not urgent. Images attached.” is searched for the word urgent, and the letter is booked as urgent by mistake. Below, three typed fields, urgency routine, images 0.97 and completeness 2 of 4, each lead to one action: a routine visit, images on file, and a request for the history. ## A type says which answers are possible Each decision gets a named slot for its answer, called a field. Each field gets a rule for what may go in it, called its type. Urgency may be one of a few words, such as *urgent*, *soon* or *routine*. The written list of fields and their types is called a schema. One way of writing schemas, JSON Schema, states the urgency rule as a list of allowed values. Any value outside the list fails the check [1]. So when a model answers "fairly urgent," the program does not have to interpret it. It rejects the answer, because "fairly urgent" is not on the list. The other two fields get their types in the same way. Whether images came is a probability, from 0 for certainly not to 1 for certainly. Completeness is a whole-number score from 1 to 4. Every answer the model can give is now either one the program was written to handle or one it can refuse. A type should also leave room for doubt, because some letters are too garbled or too short to judge. If urgency must be one of three words, the model has to pick one, and the program cannot tell a judgment from a guess. A fourth value, *unclear*, which sends the letter straight to a person, keeps that doubt visible. ## A probability lets the program set the bar For a yes-or-no question the model should give a probability, because a bare yes hides how sure the model was. With a number in hand, the program applies its own cutoff, which this note calls the bar. As an example, the clinic sets its bar at 0.9. A letter scored 0.97 for images is booked as having them, and one scored 0.55 goes to a person. If the nurse's pile then grows too large, the clinic lowers the bar. That is one number in its program, and the model need not be asked again. The bar works only if the probabilities mean what they say. Suppose the model scores a group of letters near 0.8. If about 80% of them turn out to have images, its confidence matches how often it is right. A model like that is called well calibrated [2]. Calibration cannot be assumed. A 2017 study of models that sort images and documents into classes found the newer ones poorly calibrated [2]. A 2023 study of language models found that when they are asked to say how confident they are, they tend to claim too much [3]. Neither study says how a given model will do on a given clinic's letters. So the clinic has to measure calibration itself before it trusts any bar. ## A score needs its levels written out A bare number hides a disagreement. A 3 on a four-point scale means one thing to one reader and another to the next, and a model is one more reader. So each level of the completeness score is written out in words. A 1 means the letter gives no reason for the referral, and each level above adds one thing. A 2 has a reason, a 3 adds a relevant history, and a 4 adds test results. The written levels do two jobs. They tell the model what each number means, so a score rests on something the letter contains and not on an impression. They also let a person audit the result, because anyone can read a letter scored 3 and hold it against the words for 3. If the letter has no history, the score is wrong, and nobody has to argue. ## What checking a field catches, and what it misses Because each field has a type, the program can check the whole answer before anything happens. A value outside its type is caught and the letter goes to a person, and a valid value leads to exactly one action. Some systems go a step further and hold the model to the schema while it writes. One published method steers the model's reply piece by piece, so that the reply must fit a stated pattern [4]. Its authors report that this costs little extra computing. A model steered that way cannot produce a reply that breaks the schema. Either way, the check settles the form of the answer and not its truth. A letter can come back as a perfectly valid *routine* when it should have been *urgent*. So validation catches nonsense, and only a comparison with what a careful reader would have done catches mistakes. ## Why not search a paragraph? Searching a paragraph is the simplest design to build, and that is its appeal. The model writes a short summary, and the program searches it for key words such as *urgent* or *images attached*. Such a prototype is quick to make, and the summary is useful to a person too. It breaks exactly as it did in the opening case. It can break again whenever the model's wording drifts, which may happen when the model or its instructions change. Each new phrasing is a new way for the search to misfire, and a misfire looks like any other booking. So nobody finds out until a patient who needed this week is seen next month. A check on a field behaves differently, because a value outside its type fails at once and says where. A valid but wrong value still passes in silence, as the last section said. So the fields remove one kind of silent failure, the misread wording. They leave the misjudged letter to be caught by the count of wrong actions in the check below. The paragraph's defenders are right that people need sentences. So a good design keeps one free-text field, the model's reason for its choices, written for the nurse who reviews the hard cases. The program never reads that field. People read it, and the typed fields drive the actions. ## Check it on the clinic's own letters Three checks test this answer. The first tests whether the probabilities mean what they say. Take 200 past letters that the model scored near 0.8 for images, and count how many had images. About 160 should. If 195 did, the model is underconfident and the bar can come down, and if 120 did, it is overconfident and the bar must go up. Either way the clinic sets its bar from that record and not from the model's number. The same count at other scores, such as 0.5 and 0.95, shows whether the error holds across the range. The second check tests the levels. Two staff members and the model each score the same 50 letters for completeness against the written levels. If all three give the same score on, say, 45 of the 50, the levels are clear enough to use. If they agree on far fewer, the words need sharpening before any score drives an action. The third check tests the answer itself, and it is also what catches a valid but wrong field. Run the same letters through a keyword search of the model's paragraph and through the fields. Then count the actions that differ from what a nurse reading the letter would have done. If the paragraph does as well, the fields were not worth the work. These checks leave out what each mistake costs, which differs by field. Booking an urgent patient late is worse than asking a referring office for a scan it has already sent. So the bar should be set with those costs in view. The numbers here, from the 200 letters and the 45 of 50 to the bar of 0.9, are illustrations and not data from any clinic. The form of the answer is the sure part, since a rule can check a typed value and no rule can check a sentence's meaning. Where to set the bar is not sure at all, and it has to be counted from the letters of the clinic that will use it. ## Keep Fields for code. Sentences for people. ## Sources 1. JSON Schema. Understanding JSON Schema, enumerated values: the enum keyword restricts a value to a fixed set of values. [json-schema.org](https://json-schema.org/understanding-json-schema/reference/enum) 2. Guo C, Pleiss G, Sun Y, Weinberger KQ. On calibration of modern neural networks. *ICML* 2017; arXiv:1706.04599: calibration as probability estimates that represent the true likelihood of being correct; modern neural networks, tested on image and document classification, are poorly calibrated. [arxiv.org](https://arxiv.org/abs/1706.04599) 3. Xiong M, Hu Z, Lu X, et al. Can LLMs express their uncertainty? An empirical evaluation of confidence elicitation in LLMs. arXiv:2306.13063, 2023: when verbalizing their confidence, models tend to be overconfident. [arxiv.org](https://arxiv.org/abs/2306.13063) 4. Willard BT, Louf R. Efficient guided generation for large language models. arXiv:2307.09702, 2023: guiding text generation with regular expressions and context-free grammars, with little overhead. [arxiv.org](https://arxiv.org/abs/2307.09702) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/the-need-before-the-idea/ Markdown: https://barani.org/writing/the-need-before-the-idea/index.md Published: 2026-11-05 Modified: 2026-11-05 Form: Note Topic: Mathematics teaching Series: How children learn mathematics, part 1 of 4 ======================================================================== # Why should a new idea in mathematics start with a problem a child cannot yet solve? Will the bookcase fit beside the window? A lesson in measuring starts there. Igor Barani · 5 November 2026 · 7 min read Mathematics teaching · Note · How children learn mathematics, part 1 of 4 · https://barani.org/writing/the-need-before-the-idea/ To tell which of two pencils is longer, a six-year-old lays them side by side and looks. Now ask her whether the bookcase in the hall will fit the gap between her window and the corner of her room. The method fails, because a gap cannot be carried to a bookcase, and this bookcase is too heavy to carry to the gap. At that moment she wants a better method, so the idea of measuring arrives as the answer to her own question. A new idea in mathematics should come that way, after a problem the child wants solved has defeated the method she trusts. ## A method she trusts meets a problem it cannot do The side-by-side method has served her with pencils, with ribbons and with her brother's height against her own. It has always worked, so she trusts it. The lesson depends on that trust, because only a method she relies on leaves her wanting a replacement when it fails. Here it fails at once, and she has no second method. Nobody wants to empty the bookcase and carry it in only to find it too wide. She still wants the answer. ## A string carries the comparison What she needs is something that can travel. She may think of it herself, or an adult may leave a ball of string within reach. Either way, the string now answers a question she has. A piece cut to the width of the gap can be carried to the hall and held across the front of the bookcase. If the bookcase is wider than the string is long, it will not fit. The string teaches something new. A third thing, cut to match the gap, can stand in for the gap beside the bookcase. The string is a go-between, and it carries the comparison across a distance that hands and eyes cannot cover. ## A count can be said aloud Now move the bookcase farther off. It is in a shop across town and the question comes by phone, so the string at home is no help. She needs something that can be said, and a string cannot be read aloud. A count can. Say she finds the gap nine hand-spans wide. Someone at the shop lays a hand across the bookcase again and again, counts twelve, and says it will not fit. The count has done the string's work, and done it over the phone. But was it the same hand? The count depends on two things, the length measured and the hand that measures it [1]. So nine spans of a child's hand are not nine spans of a parent's. > **Figure 1.** The gap beside the window and the bookcase, seen from above. They cannot be set side by side, and a string cut to the gap comes out short against the bookcase. A count of hand-spans answers by phone, 9 for the gap and 12 for the bookcase. The example from this note. > > What the figure shows: Three rows seen from above. First, a gap between a window and the corner of a room, and a bookcase standing apart, which cannot be set side by side. Second, a string cut to the width of the gap, held against the bookcase, comes out short. Third, the gap measures 9 hand-spans and the bookcase 12, so the bookcase will not fit. ## The names come last Only now do the ideas get their names. The hand is a unit, laying it across the gap again and again is measuring, and the count of nine is a number. Each name lands on something the child has already done and needed, so none is a bare word to memorize. The doubt about hands is the next problem, and it is already waiting. If the answer must come out the same at the shop and at home, whose hand should be used? A child who has felt that problem understands why people agreed on inches and centimeters. The ruler arrives at the end of the story, as the answer to a question she has asked. ## Where the order comes from In the Soviet Union, the psychologists D. B. Elkonin and Vasily Davydov built a school curriculum around problems like the one beside the window. They called such a problem a learning task. It is set so that children look for a general way of solving a whole class of problems [2]. The problem in front of them is only the occasion. In their system the learning task is both the basis of the general method and a source of the child's wish to learn [3]. The curriculum draws on a view of how thinking grows, in which a child develops by overcoming obstacles that her earlier methods cannot handle [4]. Classrooms have staged the same difficulty on purpose. Measure Up, an elementary mathematics project at the University of Hawaii that draws on Davydov's research, gives first graders this task [5]. Line segments are posted around the classroom to be compared with one segment, and none can be taken off the wall. So the children make a go-between of their own, a length equal to the first segment, and carry it to the others. Number arrives by the same road in Davydov's curriculum. Children first compare lengths, areas and volumes as quantities that can be equal or unequal [4]. Only then do they meet number, and they meet it through measurement [1], as how many times a chosen unit fits into a quantity [6]. Multiplication arrives that way too. Galina Zuckerman, a researcher at the Psychological Institute in Moscow, describes a lesson plan that brings second graders to it [2]. In the plan, the children can already copy a volume of water onto a second table, by counting how many times a small vessel fills. Then the only vessel the two tables share is tiny, so counting with it is possible but far too slow. The children are meant to give up measuring in one step and to measure in two. They count how many tiny vessels fill a middle-sized one, and then how many of the middle-sized ones hold all the water. At the second table they make a middle-sized measure of their own, pouring in the first count of tiny vessels and marking the level. They fill it as many times as the second count says. In one class that Zuckerman reports, the two counts were 7 and 3. Three sevens make 21, the count the small vessel alone would have needed. The teacher named what the children had done as seven times three. ## Why not start from the ruler? Handing her a ruler is quicker. Teach inches, measure the gap, and the answer is there. That order handles the measuring problems of a first-grade book, so a teacher with a crowded year has good reason to prefer it. Its weakness shows later, on a problem the ruler does not fit. Zuckerman's example, from arithmetic, is a first grader who works out how many rabbits came to visit. The same child may then struggle with the same problem about trucks bringing bricks [2]. Her general point is that solving one concrete problem correctly does not protect a child against mistakes on the next of its kind. The ruler method is exposed in the same way, because it gives the right number without ever asking why a unit is needed. A child taught the ruler first may therefore be unable to rebuild the method when the ruler is missing. ## Check it with a new problem A teacher can check a lesson plan before the lesson, and anyone can check this note's claim a week after. The plan passes if the teacher can finish the sentence "Without this idea, the child cannot yet …". If the sentence has no ending, there is no learning task yet, and the idea will arrive as a procedure. The claim is tested with a new problem of the same kind. Ask whether a rug in one room would cover a patch of floor in another, with no ruler in sight. A child who learned the general method makes a go-between or chooses a unit. A child who learned a procedure looks for the ruler, or guesses. If children taught both ways do equally well, the order of need and idea matters less than this note claims. The evidence for the order itself is thin. One study followed 46 first graders in a single Moscow school through Davydov's mathematics curriculum, with no comparison class [1]. As a group the children improved in comparison, measurement and simple addition and subtraction, however ready for school they had been at the start. But the sample was small, both teachers were new to the method, and nothing shows how the same children would have done under another curriculum. So the result is encouraging and proves nothing. Comparisons of whole curricula do exist. The author of that study cites others in which children taught in Davydov's system often show higher mathematical ability than conventionally taught children [1]. Most of those studies measured the difference at the end of fourth grade. Measure Up reports its students better prepared for algebra than their peers [6]. But such comparisons test a whole curriculum, and the order of need and idea is only one part of it. The check also leaves out the child's age, her patience, and the teacher's skill at letting a method fail without letting the child give up. A failed method can discourage a child before it interests her, so the teacher has to judge how long to let it fail. The firmest part of this note is that an idea is understood best as the answer to a question. Even that rests on a theory of learning and on classroom reports. None of the sources read for this note is a trial comparing the two orders. Whether the slower order repays its time is open. Classes taught both ways for the same hours, and then compared, would answer it. ## Keep The names come last. ## Sources 1. Sidneva A. Developmental effects of Davydov's mathematics curriculum in relation to school readiness level and teacher experience. *Frontiers in Psychology* 11:603673, 2020: number introduced through the measurement of a quantity; 46 first graders in two classes of one Moscow school, with no comparison class. [frontiersin.org](https://www.frontiersin.org/articles/10.3389/fpsyg.2020.603673/full) 2. Zuckerman G. Learning task as the heart of learning activity. *Revista Educativa* 25(1), 2022: the learning task as a search for general ways of solving a class of problems; a correct concrete solution does not protect against mistakes on the next; a lesson plan in which measuring water with a tiny vessel leads second graders to a two-step method and to multiplication. [doi:10.18224/educ.v25i1.12644](https://doi.org/10.18224/educ.v25i1.12644) 3. Elkonin BD. Modern era of the theory and practice of learning activity: key issues and perspectives. *Psychological Science and Education* 25(4):28–39, 2020: the learning task as central to the system, the basis of a general method and a source of motivation. doi:10.17759/pse.2020250403; [psyjournals.ru](https://psyjournals.ru/en/journals/pse/archive/2020_n4/Elkonin) 4. Schmittau J, Morris A. The development of algebra in the elementary mathematics curriculum of V.V. Davydov. *The Mathematics Educator* 8(1):60–87, 2004: development enabled by overcoming obstacles for which previous methods prove inadequate. [nie.edu.sg](https://math.nie.edu.sg/ame/matheduc/journal/v8_1/v81_60.aspx) 5. Dougherty B, Venenciano L. Measure Up for understanding. *Teaching Children Mathematics* 13(9):452–456, 2007: an elementary project drawing on Davydov's research; first graders compare segments posted on the wall with an intermediate measure. doi:10.5951/tcm.13.9.0452; [PDF](https://crdg.hawaii.edu/wp-content/uploads/measureup-2007-05b.pdf) 6. Venenciano L, Dougherty B. Addressing priorities for elementary school mathematics. *For the Learning of Mathematics* 34(1):18–24, 2014: in Davydov's curriculum, number is first introduced in a measurement context. [PDF](https://flm-journal.org/Articles/739D3FD8C95A0A3770B35494FA3327.pdf) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for. ======================================================================== URL: https://barani.org/writing/booked-and-used-time/ Markdown: https://barani.org/writing/booked-and-used-time/index.md Published: 2026-11-05 Modified: 2026-11-05 Form: Note Topic: Healthcare Series: Hospital finance and care, part 1 of 4 ======================================================================== # Why does a fully booked operating room stand empty? A schedule can be full while the room is not. Igor Barani · 5 November 2026 · 9 min read Healthcare · Note · Hospital finance and care, part 1 of 4 · https://barani.org/writing/booked-and-used-time/ Room 4 belongs to one surgical team every Tuesday, and this Tuesday the team has cases for the morning only. The schedule still shows the room taken all day, so it stands empty all afternoon while another team's patients wait. A booked room stands empty because the booking holds it for a team, whether or not that team has a patient for every hour. Two repairs follow, in order. Fitting each team's hours to its work removes the time that is empty every week. Opening what is still unfilled some days ahead recovers the time that is empty in some weeks. When to open it depends on how late each team books. ## A booking is a promise to a team A hospital that schedules this way divides each room's week into fixed stretches of time, such as every Tuesday in room 4. Each stretch is called a block, and it belongs to one surgeon, one group or one service such as orthopedics [1]. This note calls the holder a team, whatever its size. A block is a promise of time to a team, and the schedule records the promise. The team books its own cases into its block, and everyone else sees the room as taken. The promise buys certainty. A surgeon can plan clinic days around a known operating day, and nurses and anesthesia staff can be scheduled weeks ahead. A patient can be offered a date at the first visit. The price of that certainty can be counted. Divide the minutes a team used for its cases by the minutes its block allocated, and the result is called the block's utilization [1]. As an example, take a room open 40 hours a week and blocked to one team, whose cases fill 28 of them. The schedule shows 40 hours booked, but the room was used for 28, so its utilization is 70%. The other 12 hours stood empty with the team's name on them. > **Figure 1.** One room, one week, blocked to one team. The dotted box is time the schedule counts as booked and the room spent empty. Example numbers from this note. > > What the figure shows: Two bars for one operating room over one week. The schedule shows all 40 hours booked by one team. The room was used for 28 of them, and 12 booked hours stood empty, so the room was 70 percent used while the schedule showed it full. ## Why the block and the work drift apart A block is the same size every week, but a team's work is not. A surgeon goes to a meeting, a patient cancels, or a case planned for four hours takes two. Each leaves time inside the block with nobody to use it. The team also has a reason to hold more time than it usually needs, because time given back may not return in a busy week. A room's time cannot be stored. A drug that goes unused today stays on the shelf for tomorrow, but an hour the room stands idle on Tuesday cannot be added to Wednesday. Nobody sees the loss either, because every slot on the schedule still carries a name. ## First, fit each block to the team's work Two repairs suggest themselves. A hospital can cut each block down to the team's real work. Or it can keep the blocks and take back whatever is empty a few days ahead. The second sounds easier, since no team loses its block. The evidence is narrower than that choice, but it points one way. Once the blocks fit the work, the day on which the leftover is released hardly matters. That finding comes from an experiment on paper, published in 2004 by Franklin Dexter and Alex Macario [2]. Releasing time comes down to one decision. When a new case needs a room, whose unused time does it take? The later the decision is made, the more is known about what each team will fill for itself. So Dexter and Macario took the real schedules of two surgical suites, in which each team's time had already been allocated for efficiency. Into a day's schedule they put an invented case of one, two or three hours, and gave it to the team with the most unfilled time. They made that choice twice, once from the schedule as it stood five days before surgery and once from the schedule of the day before. Then they scored each choice by how long the rooms ran past their allocated hours. A case given to a team that then fills its own time pushes that room past its hours. They repeated this for 754 days of schedules from each suite. Deciding late ought to help, because the later schedule shows more of what each team will need, and it did. But compared with deciding five days ahead, it cut the overrun by less than 15 minutes per room per day. The authors concluded that "when OR time is released has a negligible effect on OR efficiency," where OR stands for operating room. What mattered much more, they wrote, was having the right allocations. The limits of the test follow from its design. The added cases were invented, although the schedules were real, and the suites' time was allocated for efficiency before the test began. It also scored one thing, which was overrun added up over a whole suite. So it shows that the day of release hardly changes how late the rooms run once the blocks fit. It does not show what a release rule does for blocks that are too big, and it does not show which patient gets the released hour. The next year Dexter and Richard Epstein reviewed six papers from 2004, that study among them [3]. A review is a summary and adds no second test, but it ranks the repairs the same way. When time is released mattered much less than having the right allocations in the first place. Resetting rooms faster and predicting case length better also gained little, compared with allocating time and scheduling cases well. In the example, then, the first move is to measure the team's cases and resize its block. A team whose cases fill 28 hours in a typical week does not need 40. How far above 28 to hold depends on how much its weeks vary, which the same records show. Resizing alone treats no one. Cutting the block to exactly 28 hours would lift its utilization from 70% to 100% with not one more patient seen. The gain comes only when the 12 hours go to a team with patients waiting. That is why utilization, a fair count of one block's empty time, makes a poor target. It rises whenever allocated time is cut, and it cannot tell a team that wasted its time from one that held time for urgent cases. The title of a 2024 letter by Dexter and Epstein goes further and calls utilization an invalid measure of an operating room's productivity [4]. ## Then release what is still empty, on a day that suits the team Even a block that fits will stand empty in some weeks, because no allocation can foresee a canceled case or a surgeon's week away. A release rule handles those weeks. If the team has not filled its block by a set day, the empty time opens to other teams. One company's policy, quoted in a trade journal, releases any block not scheduled 72 hours in advance, so that other physicians can book it [5]. One university hospital’s count suggests what a single deadline can do. It reviewed a year of its own cases and called a case urgent when it was booked within 48 hours of the operation [6]. Its vascular surgeons, who operate on blood vessels, were 9% of the surgeons in the study, but they booked 372 urgent cases. That was 26% of all urgent cases and more than any other team booked. The authors add that release times often fall 72 hours before the day. Set 48 hours beside 72. An urgent case is booked inside 48 hours, but the block is released at 72. So, on these numbers, a vascular block can be given away before its urgent cases exist. The time goes to a team that books ahead, and the urgent patient waits for an opening. This does not contradict the 2004 result, because the two measure different things. That experiment added up overrun across a whole suite, and a suite can finish on time while the wrong patient waits. So the experiment says the day of release barely moves how late the rooms run. The hospital's count suggests the same day can still decide who gets the hour. One hospital's count cannot give the size of that harm elsewhere, but the mechanism needs only a team that books later than the deadline. A release day should therefore follow how each team books. An executive of the company with the 72-hour policy says orthopedic and weight-loss operations are scheduled further ahead than general surgery [5]. If so, an orthopedic block still empty a week ahead will probably stay empty, and releasing it then costs little. A block for a team that books later can look just as empty on that day and still fill. Dexter and Epstein's advice fits this, since they would base any release on each team's past scheduled workload [3]. A hospital can also keep some room time unblocked for urgent cases. ## Why have a release rule at all? The 2004 experiment itself supplies the best objection. If the day of release barely changes how late rooms run, a hospital could put all its effort into allocation and have no release rule [2]. The objection is right about the size of the gain, and right that no release rule can rescue a block that is too big every week. A rule also has a cost, since schedulers must run it and teams must watch its deadline. The rule is still needed for the weeks that allocation cannot predict. A block fitted to a team's usual work will stand largely empty in the week its surgeon is away. Without a rule those hours are lost. The experiment's authors allow for this themselves, since they say how to decide "if OR time needs to be released" [2]. So each repair has its own job, and a hospital that skips the second gives up the weeks the first cannot foresee. ## Check it against last quarter's blocks Any hospital can test this answer on its own records. Take last quarter's blocks and, for each team, set the hours allocated beside the hours its cases took. Then list the hours released on each day before surgery and how many of them another team filled. Waiting times for each team belong beside those hours, because they show who paid for an empty one. The answer predicts two patterns. Most of the empty time should belong to the teams whose blocks exceed their scheduled work. Released time should be filled when it opens early enough for another team to book, and left empty when it opens too late. If instead the empty time is spread evenly across teams, or released hours fill no matter when they open, the answer is wrong for that hospital. Cancellations on the day, or slow resetting of rooms between cases, would be the next place to look. The check ignores whether a case ran longer than planned, and whether staff are free to work the hours that are moved. Either can block a reallocation that looks right on paper. It also treats one quarter as typical, which a holiday season can upset. The loss is certain, because an unused hour is gone whatever its cause. The order of the two repairs is probable, since it rests on one experiment with invented cases and one review by the same first author. The harm of an early deadline rests on one hospital's count and an inference from it. The sizes are unknown until a hospital reads its own records. ## Keep Booked is a promise. Used is a patient. ## Sources 1. Reiter K. Surgical block utilization. Impact Advisors, 7 May 2025: surgeon block utilization is the operating room minutes used for cases divided by the block minutes allocated; blocks may be assigned to a surgeon, a group or a service line. [impact-advisors.com](https://impact-advisors.com/article/surgical-block-utilization) 2. Dexter F, Macario A. When to release allocated operating room time to increase operating room efficiency. *Anesthesia & Analgesia* 98(3):758–762, 2004. Hypothetical cases added to real schedules from two surgical suites, 754 days each; release timing changed overrun time by less than 15 minutes per room per day. [doi:10.1213/01.ANE.0000100739.03919.26](https://doi.org/10.1213/01.ANE.0000100739.03919.26) 3. Dexter F, Epstein RH. Operating room efficiency and scheduling. *Current Opinion in Anaesthesiology* 18(2):195–198, 2005. A review of six papers from 2004: correct allocations matter most; release decisions should rest on historical scheduled workload. [doi:10.1097/01.aco.0000162840.02087.15](https://doi.org/10.1097/01.aco.0000162840.02087.15) 4. Dexter F, Epstein RH. Twenty-five years of previous research establish that operating room utilization is an invalid metric of theatre productivity. *JVS-Vascular Insights* 2:100118, 2024. Letter, commenting on Gupta et al. [doi:10.1016/j.jvsvi.2024.100118](https://doi.org/10.1016/j.jvsvi.2024.100118) 5. Becker's ASC Review (staff). Block time allocation best practices. 2 October 2008: one company's policy releases a block not scheduled 72 hours in advance to other physicians; it can help if a release policy recognizes that some specialties book further out. [beckersasc.com](https://www.beckersasc.com/uncategorized/block-time-allocation-best-practices.html) 6. Gupta A, Lombardi JV, Fadoul M, et al. Specialty acuity should be a consideration while assessing operating room block time metrics. *JVS-Vascular Insights* 2:100080, 2024. Single center, October 2021 to October 2022; urgent cases booked within 48 hours. doi:10.1016/j.jvsvi.2024.100080; [researchwithrowan.com](https://www.researchwithrowan.com/en/publications/specialty-acuity-should-be-a-consideration-while-assessing-operat/) --- The opinions on this site are my own and reflect my point of view, not that of the institutions I work for.