06 — Education: K-12 through College

Stage 1 — EVIDENCE BRIEF

Referee panel: Prof. Olivia Brandt (#57, statistics, C) · Prof. June Takahashi (#17, econometrics, C) · Domain specialist: Prof. Jamal Whitaker (#52, workforce education, C)

Ground rules from the panel: every number below was checked against the linked source during this round (today: 2026-09-24). Where we could not retrieve a primary source, we say "not verified this round" and nobody gets to lean on it. Effect sizes are in student-level standard deviations (SD) unless stated. Rough translation, which we will police: 0.10 SD is a meaningful but modest education effect; 0.40 SD is very large.

---

A. Achievement and the recovery

F1. NAEP 2024, grades 4 & 8 (released Jan 2025). Versus 2022: reading fell 2 points in both grades; 4th-grade math rose 2; 8th-grade math flat. All four remain below 2019. Below NAEP Basic: ~40% of 4th graders in reading (largest share since 2002), ~33% of 8th graders in reading (largest ever), ~25% 4th-grade math, ~40% 8th-grade math. NAGB: the lowest performers score "about 100 points below the highest-performing students," with the gap widening since roughly 2010. Only two state/grade/subject cells beat 2019: Louisiana 4th-grade reading, Alabama 4th-grade math. — NAGB release, 2025

F2. NAEP 2024, grade 12 (released Sept 2025). Math and reading each down 3 points vs 2019, the lowest ever reported on each assessment. 45% below Basic in math, 32% below Basic in reading (both records). At the 90th percentile scores held; lower percentiles fell, so the gap widened. More than half of seniors reported acceptance to a four-year college, a higher share than in 2019. — NAGB takeaways, 2025. Share at "college-ready" levels: 33% math (37% in 2019), 35% reading (37%). — Chalkbeat, 2025

F3. Education Recovery Scorecard (Kane/Reardon, Feb 2025; spring 2024 data). Students averaged nearly half a grade level behind 2019 in math and reading. Only 17% of grade 3–8 students are in districts above 2019 in math, 11% in reading, 6% in both. The richest-decile districts were ~4x as likely to have recovered in both as the poorest (14.1% vs 3.9%). The gap between affluent and low-income districts in math grew 11% since the pandemic. ESSER reduced losses in high-poverty districts by ~10% of a grade equivalent; districts that spent on tutoring and summer school recovered more. — Harvard CEPR, 2025

F4. Chronic absenteeism (missing ≥10% of days). ~15% in 2018-19 → 28.5% (2021-22) → 25.4% (2022-23) → 23.5% (2023-24). — AEI/Malkus, 2025. RAND estimates ~22% (~10.8M students) in 2024-25; about half of urban districts had ≥30% chronic absence; one-quarter of youth 12–21 said missing three weeks is "mostly OK." — RAND, 2025

B. Does money matter?

F5. Jackson, Johnson & Persico (QJE 2016). A 10% per-pupil spending increase sustained over all 12 school years, driven by court-ordered finance reforms: +0.27 completed years of schooling, +7.25% adult wages, −3.67 pp annual adult poverty; effects concentrated among low-income children. — NBER w20847

F6. Jackson & Mackevicius meta-analysis (AEJ: Applied 2024). $1,000 more per pupil for four years → +0.0316 SD test scores and +2.8 pp college-going; larger effects for disadvantaged students; capital and operating spending have similar marginal returns. — AEA

F7. Handel & Hanushek (Evaluation Review 2024). Using 16 test-score and 18 attainment studies: median effect of a 10% spending increase = 0.07 SD, but estimates range from −0.244 to +0.543 SD; half the test-score variance and over three-quarters of attainment variance reflect true differences across contexts that no study characteristic explains. Their warning is about generalizability, not sign. — Handel & Hanushek 2024 (PDF)_0.pdf)

F8. ESSER (CALDER WP 301, 2025). $1,000 more ESSER per pupil → +0.008 SD in district math (significant); reading similar but not significant. Full recovery at that rate would take $9,000–$13,000 more per pupil. — CALDER, 2025

C. Tutoring

F9. Saga/Match in Chicago (Guryan, Ludwig et al.). Two RCTs with 9th and 10th graders: +0.16 SD math in the first, +0.37 SD in the replication; cost $3,500–$4,300 per student per year. — NBER w28531

F10. Hybrid tech-assisted tutoring RCT (~4,000 students, 2 districts). 1:4 tutoring alternating with computer work → +0.23 SD math, about 30% cheaper than 1:2. — NBER w32510

F11. Kraft, Schueler & Falken (2024 working paper; Review of Educational Research 2026). Across 265 RCTs the pooled effect is 0.42 SD. For programs serving ≥1,000 students (US, standardized tests) it is 0.14 SD, and 0.21–0.25 SD for 400–999 students. Effects at scale are "a third to a half" of pilot effects. — EdWorkingPaper ai24-1031; J-PAL's earlier review (Nickow, Oreopoulos & Quan) is the baseline this updates: J-PAL

D. Reading and the "Southern surge"

F12. Mississippi NAEP 2024. 4th-grade reading 219 vs national 214, the first time above the national average; 4th-grade math 239 vs 237; 8th-grade reading 253 vs 257. With demographic adjustment (Urban Institute) Mississippi ranks first in 4th-grade reading and math and in 8th-grade math, and fourth in 8th-grade reading. — Mississippi First, 2025; discussion in Gelman blog, Dec 2025

F13. Mississippi's mechanism: the 2013 Literacy-Based Promotion Act. It combines coaches and phonics-focused training, K-3 screening three times a year, individual reading plans, and 3rd-grade retention (~6.5% retained in 2023). — Wikipedia summary. Fordham's 2019 retention critique noted 8% of K-3 students held back in 2018-19. Its 2022 update found the average age of 4th-grade NAEP takers was "almost identical in 2002 and 2017," so retention was not a major contributor. — Fordham

E. School choice

F14. Louisiana Scholarship Program (lottery, AEJ:Applied 2018). −0.4 SD math after one year, with negative effects in reading, science and social studies. The authors point to selection of low-quality private schools into the program. — AEA

F15. Indiana Choice Scholarship (Waddington & Berends). Public-to-private switchers: math −0.146, −0.163, −0.156 SD in years 1–3; ELA mostly null. — PMC

F16. Ohio EdChoice. Test scores: voucher users did worse than comparable public students, while eligible public schools improved modestly under competition. — Fordham/Figlio & Karbownik 2016. Attainment (Urban Institute, 2025): college enrollment 64% vs 48%, four-year 45% vs 30%, bachelor's degree 23% vs 15%, plus modest positive spillovers to eligible public schools. The design is matched comparison, not a lottery. — Urban 2025

F17. DC Opportunity Scholarship (lottery, 2010). High school graduation 82% with a voucher offer vs 70% without; math null; reading gain not significant. — Urban

F18. Florida competitive effects. Public schools facing more private-school competition improved more after the tax-credit scholarship began (Figlio & Hart, AEJ:Applied 2014). — AEA. An advocacy synthesis by the American Federation for Children reports Figlio-Hart-Karbownik (2023): ~0.166 SD in reading for high-competition public schools after 15 years. — AFC 2026. Note the source is an advocate.

F19. Arizona universal ESA. 92,362 students (Sept 15, 2025); FY2026 cost estimated at $1,036.9M against a $1,001.7M appropriation; average award $10,349. — Common Sense Institute, 2025. In the first universal year (2022-23), 71.2% of ESA students had not previously attended a public school. — LPI. EdChoice estimates net costs of the universal expansion (after switcher savings) at $178.3M in year 1 and $118.0M in year 2, i.e. 1.1% and 0.7% of AZ K-12 funding. — EdChoice 2025

F20. Federal Tax Credit Scholarship (OBBBA, P.L. 119-21). Up to $1,700 nonrefundable credit per donor starting in tax year 2027. Students qualify at ≤300% of area median income. JCT: $25.9B over 10 years. Participation is state opt-in. — BPC, ECS. 27 states had opted in by June 8, 2026. — IRS IR-2026-76

F21. Charters (CREDO NCSS3, 2023). Charter students gained +16 days of learning in reading and +6 in math per year vs matched TPS peers. CMO schools: +27 reading / +23 math; standalone charters: +10 reading / ~0 math. — CREDO press release; full report

F. Teachers, phones, early childhood

F22. Teacher pay (NEA, 2026 report). Average salary in 2024-25: $74,495 (+3.5% nominal). Average starting salary: $48,112. Average pay is ~5% lower in real terms than a decade ago. — NEA

F23. Teacher pay penalty (EPI, 2023 data). Weekly wages 26.6% below comparable college graduates (6.1% in 1996). Net of benefits the total-compensation penalty is 16.7%. — EPI

F24. Phones. 72% of high school teachers call phone distraction a major problem; 60% of those with a policy say it is hard to enforce. — Pew, 2024. Florida's statewide ban (Figlio & Özek, NBER w34388, Oct 2025): significant test-score gains in the second year and fewer unexcused absences; suspensions rose in year 1, disproportionately for Black students, then moderated. — NBER

F25. Tennessee VPK RCT (Durkin, Lipsey, Farran & Wiesen, Dev. Psych. 2022). 2,990 low-income children randomized by lottery. Children offered pre-K had lower state test scores in grades 3–6, with the largest negative effect in 6th grade, and more disciplinary infractions, worse attendance and more special-education placement. — abstract via EuropePMC, PMID 35007113

G. Higher education

F26. Sticker vs net price, 2025-26 (College Board). Published tuition and fees: public 4-year in-state $11,950; private nonprofit $45,000. Average net tuition at public 4-year in-state is ~$2,300, down from an inflation-adjusted peak of $4,450 in 2012-13. Private nonprofit net fell from $19,810 (2006-07, real) to $16,910. Grant aid has covered public 2-year tuition since 2009-10. — College Board Trends 2025

F27. Completion. Six-year completion rate for the fall-2017 entering cohort: 62.2%, flat vs the 2016 cohort. — NSC Research Center. This is the latest cohort we verified; newer cohorts were not verified this round.

F28. Student debt. $1.65T outstanding in Q2 2026; 10.6% of balances 90+ days delinquent, up from 10.3% the prior quarter. — NY Fed HHDC 2026Q2

F29. Bennett hypothesis. Lucca, Nadauld & Shen find ~60 cents of pass-through to tuition per $1 increase in the subsidized loan maximum, stronger at private and vocational programs. — NY Fed SR 733. Gordon & Hedlund's structural model attributes most of 1987–2010 tuition growth to federal loan expansion. — NBER w21967

F30. "Administrative bloat" (NCES Digest Table 314.20). Fall 1991 → fall 2022 headcounts: faculty 826,252 → 1,507,641 (+82.5%); graduate assistants +101.7%; all "other" (non-faculty) staff 1,521,232 → 1,973,819 (+29.7%). "Other" lumps growing professional and management staff together with shrinking clerical staff, and faculty counts include part-time adjuncts. — NCES

F31. OBBBA student-loan changes (final rule, Fed. Reg. May 1, 2026). Effective July 1, 2026:

• A new Repayment Assistance Plan: tiered % of AGI, $10 minimum payment, dependent adjustment.
• Grad PLUS ends for new borrowers.
• Graduate loans capped at $20,500/year and $100,000 aggregate; professional at $50,000/$200,000; lifetime cap $257,500.
• ED estimates net budget impact of −$409.3B (cohorts 1994–2035).

— Federal Register

H. Career-technical and the federal role

F32. Career Academies RCT (MDRC). Over 8 years earnings rose 11% ($2,088/yr); for young men +17% ($3,731/yr). No reduction in college-going; diploma and postsecondary credential rates were similar across groups. — MDRC

F33. Germany. About half of German school leavers enter the dual VET (apprenticeship) system. — BIBB. US Registered Apprenticeship counts and P-TECH RCT figures: we could not retrieve the DOL or MDRC pages this round, so they are not verified. No one may cite specific numbers for them.

F34. Department of Education. Headcount fell from 4,133 (Jan 2025) to ~2,183 after the March 2025 reduction in force, about half, including ~600 voluntary separations. — ED press release, Mar 2025. Later transfers of program offices to Labor, HHS and other agencies (interagency agreements, late 2025), and the Supreme Court's July 2025 stay in McMahon v. New York, are widely reported but not re-verified this round.

---

Contested evidence (where the literature genuinely disagrees)

1. Money. Jackson-Johnson-Persico and Jackson-Mackevicius find money matters, especially for poor kids. Handel-Hanushek agree the sign is usually positive but say magnitudes are unpredictable. CALDER's ESSER estimate (0.008 SD per $1,000) is roughly a quarter of the Jackson-Mackevicius per-$1,000 benchmark (0.0316 SD over four years; the time frames differ). Takahashi: the ESSER money was one-time, spent fast, and spent in a pandemic. That is weak evidence about recurring formula money, and equally weak evidence for it.
2. Tutoring at scale. Pilots show 0.3–0.4 SD; at-scale programs show 0.14 SD (Kraft et al.). Both are real. The disagreement is over whether the gap reflects fixable implementation (dosage, in-school scheduling) or regression toward the true effect.
3. Vouchers: test scores vs attainment. Test effects in the post-2012 programs (LA, IN, OH) are negative: huge in Louisiana, modest in Indiana. Attainment effects are positive or null (DC lottery grad +12 pp; Ohio's matched college results). Brandt: one of these must be mismeasured, or test scores are a poor proxy in private-school settings. We do not know which.
4. Competition effects. Figlio-Hart (Florida) and Figlio-Karbownik (Ohio) find positive spillovers to public schools. Magnitudes are modest and come from specific designs.
5. Mississippi. Real gains in grade 4; smaller in grade 8. The retention-artifact hypothesis was weakened by Fordham's age data but is not dead. No 12th-grade state NAEP exists to test persistence.
6. Pre-K. Tennessee's RCT is negative through 6th grade. Perry and Abecedarian were tiny, intensive 1960s–70s programs whose long-run gains are widely cited (specific figures not re-verified this round). Boston's lottery-based studies are reported as positive on college-going (not re-verified this round). The disagreement is about program quality vs pre-K per se.
7. Bennett hypothesis. 60-cent pass-through (Lucca et al.) vs falling real net prices at public 4-years since 2012-13 (College Board). Those can coexist, because states and institutions have raised grant aid, but advocates on each side cite only half.
8. Administrative bloat. The raw NCES headcount does not show non-faculty staff outgrowing faculty since 1991. The category is too coarse to settle whether professional and administrative staff specifically outgrew instruction.

— Brandt, Takahashi, Whitaker

Stage 2 — OPENING POSITIONS (Part 1 of 2)

---

Prof. Priya Venkataraman (#19, inequality economics, L) — "Money matters, and targeted money matters most"

Causal claim. The widening gap between the 10th and 90th percentiles (F1, F2), and the 4x recovery gap between rich and poor districts (F3), are resource problems first. Poor districts got hit harder and have less capacity to rebuild.

Evidence.

• JJP: a 10% spending increase over 12 years → +0.27 years of schooling, +7.25% wages, −3.67 pp adult poverty, concentrated among low-income kids (F5).
• Jackson-Mackevicius: $1,000/pupil over four years → +2.8 pp college-going (F6). That is not "noise."
• Even ESSER, spent in a burning building, cut losses in high-poverty districts by about 10% of a grade (F3).

Remedy. States adopt weighted formulas with a low-income weight of at least 0.3–0.5, protected by maintenance-of-effort.

Tradeoff I concede. Handel-Hanushek are right that we cannot predict the effect size in a given state (F7). Some money will be wasted on salary schedules that do not target shortages.

Falsifiable prediction. States that adopted large progressive formula changes after 2022 should show narrower low-income NAEP gaps by 2028–2030 than matched states that did not. If gaps move identically, I am wrong about marginal dollars in the current era.

---

Prof. Douglas Wren (#2, public finance, R) — "It's not the money, it's what the system does with it"

Causal claim. US real per-pupil spending rose for decades while 12th-grade scores hit record lows (F2). The binding constraint is incentive structure, not dollars.

Evidence.

• ESSER was a natural experiment of $190 billion and it produced essentially nothing.
• The median estimate in Handel-Hanushek is 0.07 SD per 10%, with a range that includes large negatives (F7). You cannot budget on a coin flip.
• CALDER: $9,000–$13,000 more per pupil to finish recovery at ESSER's rate of return (F8).

Remedy. Tie any new state money to specific evidence-based uses: tutoring, curriculum adoption, attendance teams. No general-revenue increases.

Tradeoff I concede. Court-ordered reforms of the 1970s–90s did help poor children (F5). I don't contest that history.

Falsifiable prediction. Districts that received the most ESSER per pupil will show no durable advantage on 2026 NAEP/state tests after funds expired, relative to low-ESSER districts with similar 2022 losses.

---

Prof. Kevin Marsh (#48, education policy, R) — "Choice works, and the families know it"

Causal claim. Letting families exit failing schools improves outcomes for the choosers and, through competition, for the public schools left behind.

Evidence.

• The preponderance of lottery-based evidence shows vouchers raise student achievement.
• DC lottery: graduation 82% vs 70% (F17).
• Ohio EdChoice: bachelor's degrees 23% vs 15% (F16).
• Florida competitive effects (F18).
• CREDO: +16 days reading, and much more in CMOs (F21).

Remedy. Expand means-tested ESAs; every state opts into the Federal Tax Credit Scholarship (F20).

Tradeoff I concede. Louisiana's first-year math loss is real (F14). Heavy regulation deterred good private schools from joining.

Falsifiable prediction. In FTCS opt-in states, public-school NAEP in high-competition areas should outgain non-opt-in states by 2031, and scholarship users should show higher college enrollment than matched peers.

---

Prof. Elliott Crane (#16, entrepreneurship, Lib) — "Universal ESAs are a price signal, not a subsidy"

Causal claim. Universal accounts turn a monopoly into a market. Supply of microschools, tutoring and hybrid models responds. Test scores are the wrong yardstick; parents weigh safety, values and fit.

Evidence. Arizona ESA grew to 92,362 students at an average award of $10,349 (F19). This is roughly budget-neutral: Arizona's program costs taxpayers nothing on net once you count students who leave public schools (EdChoice, F19).

Remedy. Universal ESAs with minimal testing mandates; results posted by the market, e.g. voluntary norm-referenced reporting.

Tradeoff I concede. In year one most universal users were already outside public schools (F19). That is a real transfer to families who were paying tuition.

Falsifiable prediction. Arizona's net cost per ESA user will keep falling as the switcher share rises. By FY2028 more than 60% of new users will be public-school switchers.

---

Prof. Luis Moreno (#62, sociology of work, Soc) — "Vouchers drain the commons; pay the people who do the work"

Causal claim. Universal vouchers are a regressive fiscal transfer that hollows out public schools. The teacher shortage is a wage problem.

Evidence.

• Arizona's ESA now costs about $1.4 billion a year, and 80% of recipients were already in private school.
• Louisiana −0.4 SD (F14); Indiana −0.15 SD math (F15).
• Real teacher pay fell ~5% over a decade (F22).

Remedy. Repeal universal eligibility, redirect the money to salary floors, and decline FTCS participation.

Tradeoff I concede. DC's graduation result (F17) and Ohio's college result (F16) are not nothing. Some poor families gained.

Falsifiable prediction. In universal-ESA states, public-school per-pupil state aid growth will lag non-ESA states by 2029.

Stage 2 — OPENING POSITIONS (Part 2 of 2)

---

Prof. Monica Reyes (#49, early childhood, L) — "Start at the start"

Causal claim. Gaps visible at age 9 on NAEP (F1) are largely formed before kindergarten, so high-quality pre-K is the highest-return lever.

Evidence.

• Perry and Abecedarian returned $7–$13 per dollar invested.
• Tennessee (F25) was a low-quality, fragmented program, a one-state fluke that tells us nothing about pre-K as such.

Remedy. Universal pre-K, but only with Boston-style content-rich curricula and coaching.

Tradeoff I concede. Scaled public programs have not reliably reproduced model-program results.

Falsifiable prediction. A state program meeting NIEER quality benchmarks, evaluated by lottery, should show positive 3rd-grade achievement or attainment effects. If two more quality-controlled RCTs come back null or negative, I'll concede the case rests on the 1960s studies alone.

---

Prof. Ruth Abernathy (#51, teacher education, C-L) — "The teacher pipeline is the bottleneck"

Causal claim. Tutoring, reading reform and everything else runs through teachers, and the pay gap is shrinking the pipeline.

Evidence.

• Teachers are paid 26.6% less in total than comparable graduates (EPI, F23).
• Starting pay is $48,112 (F22).
• High-dosage tutoring yields 0.37 SD when well run (F9).

Remedy. State starting-salary floors, plus differentiated pay for STEM, special ed and high-poverty schools.

Tradeoff I concede. Across-the-board raises are expensive and weakly targeted.

Falsifiable prediction. States raising floors to ≥$50k should see a smaller share of uncertified hires within three years.

---

Prof. Clara Voss (#53, neuroscience, C) — "Reading is not natural; teach it explicitly"

Causal claim. Decoding must be taught systematically. States that switched from balanced literacy to structured phonics, with screening and coaching, got results.

Evidence.

• Mississippi 219 vs 214 national in 4th-grade reading, first on adjusted ranks (F12).
• Louisiana was one of only two cells nationally to beat 2019 (F1).
• The retention critique was answered by Fordham's age data (F13), so Mississippi's gains cannot be retention artifacts.

On phones: attention is a finite resource. Florida's ban raised scores by year two (F24).

Remedy. Science-of-reading law: curriculum, training, screening, coaching. Also bell-to-bell phone bans.

Tradeoff I concede. 8th-grade gains in Mississippi are much smaller than 4th. Some of the grade-4 effect may fade.

Falsifiable prediction. States that passed science-of-reading laws in 2019–2023 should show 4th-grade reading gains relative to non-adopters on NAEP 2026/2028.

---

Prof. Yolanda Pierce-Marsh (#32, public administration, C) — "Don't break the measuring instrument"

Causal claim. Whatever one thinks of federal K-12 authority, the federal statistical and civil-rights functions (NCES/NAEP, IES evaluations, OCR) are public goods. The 2025–26 downsizing risks them.

Evidence.

• ED went from 4,133 to ~2,183 staff (F34).
• Nearly every contested finding in this thread (F1–F3, F14, F17) depends on federal data or federally funded evaluations.

Remedy. Statutorily protect NCES/NAEP and IES appropriations and staffing regardless of where K-12 program administration lands.

Tradeoff I concede. ED's grant-administration role was bloated with compliance paperwork. Moving some functions to Labor or states is defensible.

Falsifiable prediction. If capacity is hollowed out, we should see delays or cancellations of scheduled NAEP subjects and fewer IES RCTs launched in 2026–2028 vs 2015–2019.

---

Prof. Harold Finch (#50, higher-education finance, C-R) — "Cheap federal credit fed the price spiral"

Causal claim. Uncapped federal lending let institutions raise prices and grow non-instructional staff. Borrowers carry the result: $1.65T in debt, 10.6% seriously delinquent (F28).

Evidence.

• Federal loan increases pass through to tuition essentially dollar for dollar (Lucca et al., F29).
• Administrative staff have grown far faster than faculty for 30 years.

Remedy. Keep the OBBBA loan caps and Grad PLUS elimination (F31). Add institutional risk-sharing on unpaid balances.

Tradeoff I concede. Caps may push some graduate students into costlier private loans.

Falsifiable prediction. Graduate programs that relied most on Grad PLUS should show slower tuition growth in 2026-27 through 2028-29 than programs that didn't.

Stage 3 — CROSS-EXAMINATION (Part 1)

Exchange A: Venkataraman (L) ⟷ Wren (R) on money (4 rounds)

Venkataraman → Wren, R1. Douglas, "ESSER produced essentially nothing" is not what the evidence says. CALDER finds a statistically significant +0.008 SD in math per $1,000 (F8). Kane/Reardon find ESSER cut high-poverty losses by ~10% of a grade (F3). Small is not zero. And your own concession, that court-ordered reforms worked (F5), undercuts "it's not the money."

Wren, R1 response. I'll amend "essentially nothing" to "a small positive effect at an enormous price." Now scale your own numbers. At 0.008 SD per $1,000, the ~$190B, which is on the order of $3,000–4,000 per pupil, bought roughly 0.03 SD. The loss was about half a grade (F3). The CALDER authors say full recovery would take $9–13k more per pupil. JJP's money was sustained for twelve years and flowed through stable systems. ESSER was a sugar high: much of it went to HVAC, one-time bonuses, and hiring that was reversed at the 2024 fiscal cliff. If your mechanism is "sustained, targeted funding," then ESSER is not your evidence and my critique stands against untargeted money.

Venkataraman, R2. Agreed on the mechanism, and that is my point. Jackson-Mackevicius pool dozens of sustained spending changes and get +0.0316 SD and +2.8 pp college-going per $1,000 over four years (F6). They find minimal publication bias and similar returns for capital and non-capital spending. That is four times the ESSER per-dollar yield. The comparison you want is formula money, not a one-time pandemic transfer. Hanushek's own 2024 paper puts the median at +0.07 SD per 10% (F7). You're citing the range to imply the center is zero.

Wren, R2. I'm citing the range because Handel-Hanushek say explicitly that we can't tell which part of the distribution a new program will land in (F7), and three-quarters of the attainment variance is unexplained true heterogeneity. A legislature doesn't get the median; it gets a draw. So conditional money makes sense: tie dollars to the uses Kane/Reardon found effective (tutoring, summer school, F3).

Venkataraman, R3. Then we have converged on something testable: dollars plus use restrictions. But note Jackson-Mackevicius found capital and operating dollars have similar marginal returns (F6). That is inconvenient for anyone, you included, who thinks the use category is what matters. The effect might come from sustained resources in poor districts, whatever they're spent on.

Wren, R3. Or from something unmeasured about which states reform finance and when. That's Handel-Hanushek's point about "contexts of convenience." I concede the sign; I don't concede that the next dollar in Illinois looks like the average dollar in a 1990s court order. I'll accept a formula weight if it comes with outcome reporting by subgroup. That is movement, not a cave.

Takahashi (referee note). Both are now consistent with F5–F8. The residual disagreement is external validity, which the evidence cannot currently resolve.

---

Exchange B: Abernathy (C-L) ⟷ Marsh (R) on vouchers (3 rounds)

Abernathy → Marsh, R1. Kevin, "the preponderance of lottery-based evidence shows vouchers raise achievement" is false for the programs that look like what you're proposing. Louisiana's lottery: −0.4 SD math (F14). Indiana: −0.15 SD math for switchers, all three years (F15). Ohio: voucher users did worse (F16). These are the modern, statewide programs. The 1990s positive estimates came from small urban programs in a very different private-school market.

Marsh, R1 response. I'll narrow the claim. On test scores the modern evidence is negative or null, and I withdraw "preponderance." But look at attainment, which is what JJP and Venkataraman say matters. DC's lottery: +12 pp graduation (F17). Ohio: college enrollment 64% vs 48%, BA 23% vs 15% (F16). The Ohio study is from Urban, not a choice advocate. If test scores and life outcomes disagree, I'll take life outcomes.

Abernathy, R2. The Ohio attainment study is a matched comparison, not a lottery (F16). Families who use vouchers differ in ways test-score matching won't catch; Figlio and Karbownik themselves found participants were less disadvantaged than eligible non-users. DC is a real lottery, but it was a small program in 2004–2009 in one city. You cannot build universal ESAs on it.

Marsh, R2. Selection cuts both ways. Figlio-Karbownik also found participants in Ohio had lower test scores than the public students they were compared with, yet went on to more college. And competitive effects don't have a selection problem in the same way: Figlio-Hart in Florida (F18) and Figlio-Karbownik in Ohio (F16) both found public schools improved when facing competition.

Abernathy, R3. Modestly, yes, and I'll grant it. But Louisiana's −0.4 SD is roughly the size of the entire pandemic loss in reading. A policy that risks that for switchers while producing modest spillovers needs quality screens on participating schools, and your universal-ESA allies oppose those.

Marsh, R3. On that we agree more than you'd think. Louisiana's regulatory design pushed good schools out (F14 authors' own interpretation). I'll support mandatory norm-referenced testing of voucher users, with results published by school. That is a concession Crane won't like.

---

Exchange C: Moreno (Soc) ⟷ Crane (Lib) on Arizona (3 rounds)

Crane → Moreno, R1. Luis, "$1.4 billion" and "80% already private" are wrong. The FY2026 estimated cost is $1,036.9M (CSI, F19). LPI's own number for 2022-23 is 71.2% not previously in public school, and it wasn't "80% private": that group includes kindergartners and homeschoolers. CSI reports that over half of new enrollees now come from public schools.

Moreno, R1 response. I'll correct the figures: ~$1.04B, 71.2% in year one. The point survives. Most of the first-year universal money went to families who were not in public schools, a new entitlement costing hundreds of millions. And your "costs taxpayers nothing" claim is contradicted by your own source: EdChoice estimates net costs of $178M and $118M in years 1–2 (F19).

Crane, R2. Fair. I retract "nothing." My claim is that net costs are small, 0.7–1.1% of Arizona K-12 funding by EdChoice's estimate, and falling as the switcher share rises. That's my falsifiable prediction.

Moreno, R3. EdChoice's savings assume that every "switcher" would otherwise have stayed in public school at full state cost, and that districts can shed costs proportionally. Neither is guaranteed. Fixed costs don't leave with the kid. I'll hold my line that it's a transfer; I accept that the size is disputed.

---

Exchange D: Stroud (R, #63, family sociology) → Reyes (L) on pre-K

Stroud. Monica, you call Tennessee a "fluke." It's the only large-scale RCT of a statewide program, with 2,990 kids (F25). The children offered pre-K did worse by 6th grade on tests, discipline, attendance and special-ed placement. And "$7–$13 per dollar" is a number I've never seen reconciled across Perry and Abecedarian. What's your source?

Reyes. I can't produce a verified source for the $7–$13 figure this round; the referees flagged Perry/Abecedarian numbers as not re-verified (Brief, Contested #6). I withdraw the specific ratio. On Tennessee, I withdraw "fluke." It is strong evidence that scaled, low-quality pre-K can harm. My claim is narrowed: pre-K quality is the treatment. I would bet on Boston-type programs. But Boston's results weren't verified here either, so I hold that as a prediction, not a finding.

Stroud. Then the burden sits where it should. Show a quality-controlled statewide RCT that's positive before we fund universal pre-K.

Reyes. Agreed on the RCT demand. Not agreed that we halt existing programs while waiting.

Stage 3 — CROSS-EXAMINATION (Part 2)

Exchange E: Oyelaran (C-R, #12, finance) → Abernathy (C-L) on teacher pay

Oyelaran. Ruth, EPI's 26.6% is the weekly wage gap. Their own total-compensation gap, after benefits, is 16.7% (F23). You said "26.6% less in total," which is wrong. Separately, 0.37 SD is Saga's replication estimate; the first Chicago trial got 0.16 (F9), and at scale the literature says 0.14 (F11).

Abernathy. Corrected on both. The compensation penalty is 16.7%, still a record in EPI's series. Saga's two trials were 0.16 and 0.37. My larger point is untouched: NEA's data show real average pay down ~5% over a decade (F22), during a period when district spending rose. Something other than teacher salaries absorbed the money, which Wren should find interesting.

Oyelaran. Agreed that's a real puzzle. Part of the answer is benefits and pension costs, which EPI's 9.9% benefit advantage reflects. That's compensation teachers receive, even if it's back-loaded and poorly valued by the young teachers you're trying to recruit.

Abernathy. Which is an argument for restructuring compensation toward front-loaded pay. I'll take that.

---

Exchange F: Walker (L, #65) → Voss (C) on Mississippi

Walker. Clara, you said Mississippi's gains "cannot" be retention artifacts. Fordham's age check (F13) compares average age in 2002 and 2017. Averages can hide compositional change among the low scorers, exactly the kids retention targets. And Mississippi's 8th-grade reading is still 4 points below the nation (F12).

Voss. "Cannot" was too strong; I amend it to "the evidence doesn't support retention being the main driver." The age data rule out large compositional effects, not small ones. But look at the pattern: the demographic-adjusted #1 rank holds in 4th-grade math too (F12), and math is untouched by a reading-based retention rule except through age. Louisiana, without Mississippi's retention intensity, also beat 2019 in 4th-grade reading (F1).

Walker. Math could still reflect age if retained kids are older. And Louisiana's 4th-grade reading has its own policy story.

Voss. Then the test is the one I proposed: state NAEP 2026/2028 for science-of-reading adopters vs non-adopters, stratified by whether they include mandatory retention. If gains appear only where retention is mandatory, you win.

---

Exchange G: Vogt (C-L, #11) → Finch (C-R) on Bennett and bloat

Vogt. Harold, two overstatements. First, Lucca-Nadauld-Shen found ~60 cents on the dollar, and specifically for subsidized loan maximums, with smaller effects for unsubsidized loans (F29). Not "dollar for dollar." Second, average net tuition at public four-years fell from $4,450 (real, 2012-13) to ~$2,300 in 2025-26 (F26), while federal lending continued. And your bloat claim: the NCES headcount shows faculty +82.5% and all non-faculty staff +29.7% since 1991 (F30).

Finch. I amend to "around 60 cents for subsidized loans." Gordon-Hedlund's model goes further (F29), but a model isn't an identification strategy. On net price, sticker prices still rose and aid rose faster, so the falling net price is partly institutional discounting financed by... higher sticker prices. The Bennett story lives in graduate and professional programs, which is where Grad PLUS was uncapped.

On bloat, I withdraw "administrators far faster than faculty" as a claim supported by that table. The NCES "other" category lumps clerical jobs, which have shrunk, with professional staff, and faculty growth is heavily part-time. The table can't prove me right. It also doesn't prove me wrong, so I'll call it unsupported pending IPEDS occupation-level data.

Vogt. Then we agree the grad-program version is the live hypothesis, and OBBBA's caps (F31) are a natural experiment for it.

---

Exchange H: Ashford (R, #31, federalism) → Pierce-Marsh (C) on ED

Ashford. Yolanda, the 12th-grade NAEP was released on schedule in September 2025, after the RIF (F2, F34). Where's the damage?

Pierce-Marsh. Fair point. That release proves the core schedule survived its first year. It doesn't prove IES evaluation capacity survived, and I have no verified count of cancelled studies this round. I'll hold my claim as a prediction (fewer IES RCTs 2026–28), not a finding.

Ashford. Then I'll support protecting NAEP and NCES by statute, as long as that isn't a back door to keeping program offices.

---

Exchange I: Marsh (R) → Moreno (Soc) on competition

Marsh. Luis, your "drain the commons" model predicts public schools get worse under choice. Florida (F18) and Ohio (F16) show they got modestly better, with Ohio's spillover including higher college attendance in eligible public schools.

Moreno. Those are means-tested programs at modest scale. My fiscal claim is about universal programs in which the marginal user was already private. Competition effects from a 5% exit don't tell you what 15% does to fixed-cost districts. I'll concede the modest spillover in targeted programs.

---

Exchange J: Reyes (L) → Voss (C) on phones

Reyes. Figlio-Özek found suspensions rose in year one, disproportionately for Black students (F24).

Voss. Yes, and it's the implementation cost I'd flag in any bill. Enforcement design matters: storage pouches or lockers rather than confiscation-and-discipline. The year-two gains and fewer unexcused absences are the benefit side.

Stage 4 — REFEREE FACT-CHECK

Brandt (#57), Takahashi (#17), Whitaker (#52)

We rate every empirical claim made in Stages 2–3, as originally stated. Where a professor amended during cross-examination we note it, but the rating applies to the original.

# · Claimant · Claim · Rating · Basis
1 · Venkataraman · 10% more spending for 12 yrs → +0.27 yrs schooling, +7.25% wages, −3.67pp poverty; larger for low-income · Supported · F5 (JJP)
2 · Venkataraman · $1,000/pupil for 4 yrs → +2.8pp college-going · Supported · F6
3 · Venkataraman · ESSER cut high-poverty losses ~10% of a grade · Supported · F3 (ERS)
4 · Venkataraman · Rich/poor district recovery gap ~4x · Supported · F3: 14.1% vs 3.9%
5 · Venkataraman · Jackson-Mackevicius per-$ yield ~4x ESSER's · Contested · 0.0316 SD over 4 yrs vs 0.008 SD; different time frames and outcome levels (student vs district). The direction is fair, the multiple is loose
6 · Venkataraman · Capital and operating spending have similar returns · Supported · F6
7 · Wren · ESSER "produced essentially nothing" · Wrong · F8: significant +0.008 SD/$1,000 in math; F3: ~10% grade reduction in high-poverty losses. Amended on record
8 · Wren · ESSER ≈ $190B, ≈ $3–4k/pupil, bought ~0.03 SD · Contested · Total and per-pupil figures not verified in our brief; the arithmetic follows from F8 only if the full amount is applied to CALDER's ESSER-III-based estimate
9 · Wren · ESSER largely spent on HVAC, one-time bonuses, reversed hiring · Unsupported · No spending-composition source verified this round
10 · Wren · Median 0.07 SD per 10%; range includes negatives; most variance unexplained · Supported · F7
11 · Wren · Full recovery would need $9–13k/pupil at ESSER rates · Supported · F8
12 · Wren · 12th-grade scores at record lows · Supported · F2
13 · Marsh · "Preponderance of lottery evidence shows vouchers raise achievement" · Wrong · F14, F15, F16: modern statewide programs show negative or null test effects. Retracted on record
14 · Marsh · DC graduation 82% vs 70% · Supported · F17 (lottery)
15 · Marsh · Ohio BA 23% vs 15%; college 64% vs 48% · Supported (as figures) / Contested (as causal) · F16. Matched design, not a lottery
16 · Marsh · Florida/Ohio competition improved public schools · Supported (modest) · F16, F18. The 0.166 SD figure comes via an advocacy summary
17 · Marsh · CREDO +16 reading / CMOs much more · Supported · F21
18 · Marsh · Louisiana's losses due to regulation deterring good schools · Contested · The authors cite low-quality school selection; the "regulation caused it" mechanism is an inference, not their finding
19 · Crane · Arizona 92,362 students, $10,349 average award · Supported · F19
20 · Crane · Arizona ESA "costs taxpayers nothing on net" · Wrong · F19: even EdChoice estimates net costs of $178M / $118M. Retracted on record
21 · Crane · Net cost ~0.7–1.1% of AZ K-12 funding · Supported (as EdChoice's estimate) / Contested (as fact) · F19. Depends on the switcher counterfactual; Moreno's fixed-cost objection is valid
22 · Crane · Over half of new enrollees now come from public schools · Supported · F19 (CSI)
23 · Moreno · Arizona ESA costs ~$1.4B/yr · Wrong · F19: FY2026 est. $1,036.9M. Amended
24 · Moreno · 80% of recipients already in private school · Wrong · F19: 71.2% not previously in public school in 2022-23, a group that includes kindergartners and homeschoolers. Amended
25 · Moreno · LA −0.4 SD; IN −0.15 SD math · Supported · F14, F15
26 · Moreno · Real teacher pay −5% over a decade · Supported · F22
27 · Reyes · Perry/Abecedarian return $7–$13 per $1 · Unsupported · Not verified this round. Withdrawn on record
28 · Reyes · Tennessee a "one-state fluke" · Wrong (as framed) · F25 is a large RCT; a single study, but not a fluke. Withdrawn on record
29 · Stroud · TN pre-K: worse tests, discipline, attendance, special ed by grade 6 · Supported · F25
30 · Abernathy · Teachers paid "26.6% less in total" · Wrong · F23: 26.6% is weekly wage; total compensation gap is 16.7%. Amended
31 · Abernathy · Starting pay $48,112 · Supported · F22
32 · Abernathy · Tutoring yields 0.37 SD when well run · Contested · F9: 0.16 and 0.37 across two Saga RCTs; F11: 0.14 at scale
33 · Oyelaran · Benefits advantage 9.9% · Supported · F23
34 · Voss · Mississippi 219 vs 214; #1 adjusted · Supported · F12
35 · Voss · Mississippi gains "cannot be" retention artifacts · Contested · F13 weakens the artifact story; averages can mask composition. Amended
36 · Voss · Louisiana beat 2019 in 4th-grade reading · Supported · F1
37 · Voss · Florida phone ban raised scores by year two · Supported · F24 (NBER WP, not yet peer reviewed)
38 · Walker · Suspensions rose, disproportionately for Black students · Supported · F24
39 · Walker · Mississippi 8th-grade reading 4 points below nation · Supported · F12
40 · Pierce-Marsh · ED staff 4,133 → ~2,183 · Supported · F34
41 · Pierce-Marsh · IES evaluation capacity damaged · Unsupported · No verified data; reframed as prediction
42 · Ashford · 12th-grade NAEP released on schedule after the RIF · Supported · F2 (Sept 2025), F34
43 · Finch · $1.65T debt, 10.6% seriously delinquent · Supported · F28
44 · Finch · Loans pass through "essentially dollar for dollar" · Wrong · F29: ~60¢, for subsidized maximums. Amended
45 · Finch · Admin staff grew far faster than faculty for 30 yrs · Unsupported · F30 headcount shows the opposite at the coarse level; occupation detail not verified. Withdrawn
46 · Finch · Falling net price partly financed by high-sticker discounting · Contested · Plausible; no source verified this round
47 · Vogt · Public 4-year net tuition $4,450 → ~$2,300 (real) · Supported · F26
48 · Crane / Moreno · Most year-1 universal users were not public-school switchers · Supported · F19

Tally (48 claims)

• Supported: 31 (the three split ratings, #15, #16 and #21, are counted as Supported here)
• Contested: 7 (#5, #8, #18, #32, #35, #46, plus the causal half of #15, counted once)
• Unsupported: 4 (#9, #27, #41, #45)
• Wrong: 8 (#7, #13, #20, #23, #24, #28, #30, #44)

Corrected count by line: Supported = 1,2,3,4,6,10,11,12,14,15,16,17,19,21,22,25,26,29,31,33,34,36,37,38,39,40,42,43,47,48 = 30. Contested = 5,8,18,32,35,46 = 6. Unsupported = 9,27,41,45 = 4. Wrong = 7,13,20,23,24,28,30,44 = 8. Total 48. These line counts are the official tally.

Pattern note (Brandt)

Overstatement was bipartisan. Every "Wrong" rating went to a professor whose direction was defensible but who inflated the magnitude or the certainty:

• Right: "ESSER nothing," "preponderance," "costs nothing," "dollar for dollar."
• Left: "$1.4B / 80%," "fluke," "26.6% total."

All eight were retracted or amended on the record without the referees having to press twice.

Stage 5 — STEELMAN

Left steelmans Right's case on school choice

Written by Venkataraman (#19, L)

The strongest case for choice doesn't rest on test scores. It rests on four points.

1. The attainment evidence is better than the test evidence, and attainment is what JJP (F5) taught us to care about. DC's lottery (+12 pp graduation, F17) is as clean a design as anything in this thread. Ohio's college effects (F16) are large, concentrated among the lowest-income and lowest-scoring students, and replicate the pattern.
2. Competition effects are consistently positive in Florida and Ohio (F16, F18), and public-school students are the vast majority.
3. Money without exit hasn't reliably worked. Handel-Hanushek (F7) show the payoff to new spending is unpredictable. ESSER's per-dollar yield was small (F8). Twelfth-grade scores are at record lows (F2) despite rising real spending.
4. Families value things tests don't measure (safety, values, fit), and a liberal society should respect that.

The Louisiana result (F14) is a warning about program design, not a refutation of choice.

Right's reply — Marsh (#48, R): Accepted as fair. I'd add only that CREDO's CMO results (F21) show that choice within the public sector also works when quality operators can scale. — Crane (#16, Lib): Accepted with one correction. Point 4 is not a side argument for libertarians; it's the main argument. Tests are the yardstick of the system we're trying to exit.

---

Right steelmans Left's case on public investment

Written by Wren (#2, R)

The strongest case goes like this:

1. The best-identified long-run evidence we have, from court-ordered finance reforms (F5) and a meta-analysis with minimal publication bias (F6), says sustained money raises attainment and earnings for poor children. Those are exactly the outcomes Marsh wants to judge vouchers by.
2. The post-pandemic gap is a poverty gap (F3: 14.1% vs 3.9% recovery). Absenteeism concentrates in urban districts (F4), and teacher pay has fallen in real terms (F22, F23). Those are resource-sensitive problems.
3. The modern voucher programs produced some of the largest negative test effects in the education literature (F14, F15). Arizona's universal program mostly subsidized families who were not in public schools in year one (F19).
4. Rejecting money because ESSER was spent badly is like rejecting medicine because one emergency dose was badly timed.

Left's reply — Venkataraman: Accepted as fair. Moreno (#62, Soc): Accepted, with a correction. It understates the collective-goods argument. Public schools are a democratic institution, not just a production function. A choice regime that fragments the common school has costs no NAEP score captures.

---

Center steelmans both on reading and phones

Written by Takahashi (#17), as referee, at the request of Voss and Walker

Pro (Voss). Mississippi's adjusted-#1 results in both 4th-grade reading and math (F12), Louisiana's lone beat of 2019 (F1), and the failure of the retention-artifact hypothesis on age data (F13) together make the science-of-reading package the most promising scalable K-3 reform in a decade.

Skeptic (Walker). The effect is concentrated at grade 4 and attenuates by grade 8 (F12). Mandatory retention is bundled into the package and its separate effect is unidentified. Nobody has a clean causal design, only state-level before-and-after comparisons.

Both reply: Accepted as fair.

---

Center steelmans Right on higher ed

Written by Vogt (#11, C-L)

Federal credit that is uncapped at the graduate level plausibly lets programs raise prices. Lucca et al. (F29) find 60¢ pass-through on subsidized limits, with larger effects at private and vocational programs. $1.65T outstanding and 10.6% serious delinquency (F28) show many borrowers didn't get value.

Finch's reply: Accepted as fair, and more honest than my opening.

Stage 6 — REVISED PROPOSALS

Drafted by the referee panel from the positions as amended in Stages 3–5.

---

P1. Evidence-conditioned high-dosage tutoring

Mechanism. States fund in-school, scheduled tutoring (≥3 sessions/week, groups of ≤4, a consistent tutor), including the hybrid human + software model. Every district program above 1,000 students must run a lottery or phase-in evaluation and publish effects.

Cost. Saga: $3,500–$4,300 per student per year (F9). The hybrid model is ~30% cheaper (F10). Covering 10% of grade 6–9 students in a mid-size state is on the order of hundreds of millions per year. No official score.

Precedent. Chicago's two Saga RCTs: 0.16 and 0.37 SD (F9). Large-scale programs nationally average 0.14 SD (F11), which should be the budgeting assumption.

Key risk. Dosage erosion at scale: tutoring gets scheduled outside school hours or staffed with undertrained tutors.

---

P2. Science-of-reading package, without mandatory retention

Mechanism.

• Adopted curricula from state-approved structured-literacy lists.
• K-3 screening three times a year.
• Individual reading plans.
• State-funded literacy coaches in low-performing schools.
• Teacher-prep licensure tests aligned to structured literacy.

Cost. No official score. Mississippi's coaching appropriation was not verified this round.

Precedent. Mississippi 2013 LBPA: 4th-grade reading from below to above the national average by 2024 (F12, F13). Louisiana beat 2019 in 4th-grade reading (F1).

Key risk. Curriculum adoption without training fidelity. Gains may concentrate at grade 4 and fade (F12, 8th grade).

---

P3. Mandatory 3rd-grade retention with intensive intervention (Mississippi-style)

Mechanism. Students below the reading cut score after 3rd grade repeat the grade with a highly rated teacher and daily intervention, with good-cause exemptions.

Cost. One additional year of schooling per retained student, roughly the state's per-pupil cost × 6–8% of a cohort (F13 retention rates). No official score.

Precedent. Mississippi (F13), Florida.

Key risk. Stigma and dropout effects. The separate causal effect within the Mississippi bundle is unidentified.

---

P4. Bell-to-bell phone restrictions with non-punitive enforcement

Mechanism. State law requires phones to be stored (pouches or lockers) for the whole school day. Enforcement is confiscation-first, not suspension. States must publish discipline data by race in years 1–2.

Cost. Low. Pouches run a few dollars to ~$30 per student (vendor pricing, not verified). No official score.

Precedent. Florida: test gains and fewer unexcused absences in year two; a year-one suspension spike that fell disproportionately on Black students (F24).

Key risk. Discipline disparities in year one (F24). Pew: 60% of HS teachers with policies find them hard to enforce.

---

P5. Accountability for publicly funded private-school choice

Mechanism. Any voucher, ESA or FTCS-funded student takes a nationally norm-referenced test annually. Results are published by school when n≥10. Funding flows are audited and published. Schools with sustained large negative value-added lose eligibility.

Cost. Testing ~$20–$50 per student (estimate, not verified). No official score. FTCS itself: $25.9B over 10 years (JCT, F20).

Precedent. Louisiana required state testing, which is how F14 was detected. Indiana's data allowed F15.

Key risk. Private-school supply shrinks, the mechanism Marsh blames for Louisiana (Contested, FC #18). Libertarians argue it standardizes curricula.

---

P6. Progressive state funding weights with maintenance-of-effort

Mechanism. Low-income weight of ≥0.3 in state formulas, no cuts to per-pupil base in real terms, and public subgroup outcome reporting (Wren's condition).

Cost. Jackson-Mackevicius benchmark: $1,000/pupil for 4 years → +0.0316 SD, +2.8 pp college (F6). Weighted increases for ~half of students imply roughly 3–5% of state K-12 budgets. No official score.

Precedent. School-finance reforms 1970s–2000s (F5).

Key risk. Handel-Hanushek heterogeneity (F7): the next state may land low in the distribution.

---

P7. Higher-ed institutional risk-sharing plus price transparency

Mechanism. Institutions reimburse a share (e.g., 15–20%) of federal loan balances unpaid after X years, by program. Mandatory net-price disclosure by income band on award letters. This builds on the OBBBA caps and RAP (F31).

Cost. Would raise federal revenue relative to baseline. No official score for this design. ED scores the OBBBA loan provisions at −$409.3B (F31).

Precedent. None at federal scale. Graduate-program exposure to Grad PLUS ending (F31) will provide a natural experiment.

Key risk. Institutions reduce access to riskier (lower-income) students.

---

P8. Scale Career Academies and registered apprenticeship with rigorous evaluation

Mechanism. Federal and state matching grants for career academies (school-within-a-school plus employer partnerships and work-based learning). Expanded youth registered apprenticeships. Every new cohort evaluated by lottery where oversubscribed.

Cost. No official score.

Precedent. Career Academies RCT: +11% earnings over 8 years, +17% for young men, no loss in postsecondary credentials (F32). In Germany about half of school leavers enter dual VET (F33). The institutional context differs sharply from the US.

Key risk. Employer participation in the US is thin (DOL counts not verified this round). Sites may cream the most employable students.

Stage 7 — RECORDED VOTE

Blocs: Left 34 · Center 33 · Right 32. Moderator #100 abstains.
Bridge rule: consensus requires ≥60% in every bloc (Left ≥21, Center ≥20, Right ≥20).

# · Proposal · Left yes · Center yes · Right yes · Consensus?
P1 · Evidence-conditioned high-dosage tutoring · 27/34 (79%) · 28/33 (85%) · 21/32 (66%) · YES
P2 · Science-of-reading package (no mandatory retention) · 25/34 (74%) · 29/33 (88%) · 23/32 (72%) · YES
P3 · Mandatory 3rd-grade retention + intervention · 10/34 (29%) · 20/33 (61%) · 28/32 (88%) · NO
P4 · Bell-to-bell phone restrictions, non-punitive enforcement · 22/34 (65%) · 27/33 (82%) · 27/32 (84%) · YES
P5 · Testing and transparency for publicly funded private choice · 31/34 (91%) · 29/33 (88%) · 16/32 (50%) · NO
P6 · Progressive funding weights + maintenance of effort · 32/34 (94%) · 23/33 (70%) · 11/32 (34%) · NO
P7 · Higher-ed risk-sharing + net-price transparency · 24/34 (71%) · 27/33 (82%) · 22/32 (69%) · YES
P8 · Career academies / apprenticeship scale-up with RCTs · 26/34 (76%) · 28/33 (85%) · 23/32 (72%) · YES

Five consensus reforms: P1, P2, P4, P7, P8.

On P2, P1 and P8 the Right split along libertarian lines: the "no" votes object to state-mandated curriculum lists and new grant programs, not to the evidence.

---

Vote explanations

Prof. Kevin Marsh (#48, R): YES on P5. One of 16 Right yeses.
"I voted with the Left on this, and against most of my bloc. After the cross-exam I can't defend −0.4 SD (F14) as a design choice. Testing is how we learned about it.
What would change my mind: evidence that testing mandates themselves caused the Louisiana supply collapse, i.e. a design comparing states with and without mandates showing a large drop in high-quality school participation."

Prof. Elliott Crane (#16, Lib): NO on P5, NO on P6, YES on P4 (reluctantly).
"P5 standardizes the thing families are exiting. P6 bets on the median of a distribution Handel-Hanushek say we can't locate (F7).
What would change my mind on P5: evidence that universal ESA users in Arizona show large negative outcomes on attainment, not just tests. Then accountability would be protecting children rather than protecting the system."

Prof. Luis Moreno (#62, Soc): NO on P3, NO on P7, YES on P5 and P6.
"On P7, risk-sharing will make colleges shun poor students. On P3, retention is punishment dressed as rigor.
What would change my mind on P3: a credible design, such as a regression discontinuity at the cut score, showing retained students with higher graduation rates and no rise in dropout."

Prof. Priya Venkataraman (#19, L): YES on P6, YES on P1, NO on P3.
"The consensus list is good but tells you where the Right will spend, not how much.
What would change my mind on P6: a well-identified post-2020 study of a sustained formula increase showing null attainment effects. That would move me toward Hanushek's generalizability critique."

Prof. Douglas Wren (#2, R): NO on P6, YES on P1, YES on P7.
"I moved from 'it's not the money' to 'it's conditioned money.' P6 lacks the conditions.
What would change my mind: if states that adopted weights after 2022 show gap narrowing on NAEP 2028 relative to matched states (Venkataraman's own prediction). That would move me to yes."

Prof. Clara Voss (#53, C): YES on P2, YES on P4, YES on P3 (narrowly).
"I'll say plainly that my yes on P3 rests on weaker evidence than my yes on P2.
What would change my mind on P3: NAEP 2026/2028 showing adopters without retention gaining as much as adopters with it (Walker's test from Exchange F)."

Stage 8 — VERDICT

Prof. Adelaide Wainwright (#100, moderator, institutional design, C). I did not vote.

Established (strong evidence, cross-bloc agreement after fact-check)

• The decline is real and it is bottom-heavy.
• Every NAEP grade and subject is below 2019.
• 12th grade hit record lows in 2024.
• The 90th percentile held while lower percentiles fell (F1, F2).
• Recovery tracks district income by roughly 4x (F3).
• Chronic absenteeism has come down from 28.5% but remains ~22–23.5%, about 1.5x pre-pandemic (F4).
• Money matters on average, especially sustained money for poor children (F5, F6). Even Hanushek's meta-analysis puts the median positive (F7).
• ESSER had small positive effects at high cost (F3, F8). Neither "it did nothing" nor "it proves money works" survived.
• Tutoring works, but at scale roughly a third as well as in pilots (0.42 → 0.14 SD, F11).
• Modern statewide voucher programs have produced negative test-score effects, large in Louisiana and modest in Indiana and Ohio (F14–F16).
• Tennessee's statewide pre-K RCT was negative through 6th grade (F25).
• Real net tuition at public four-years has fallen since 2012-13 even as sticker prices rose (F26).
• Teachers face a record wage penalty. 26.6% weekly, 16.7% in total compensation (F23).

Contested (credible evidence on more than one side)

• Voucher attainment effects (DC lottery positive; Ohio positive but non-experimental) vs test effects. The reason for the divergence is unknown.
• Size and sustainability of competition spillovers (F16, F18).
• Mississippi's gains: instruction vs retention vs composition, and whether grade-4 gains persist (F12, F13).
• Arizona ESA net cost: it depends entirely on the switcher counterfactual (F19).
• The Bennett hypothesis: ~60¢ pass-through on subsidized limits (F29) vs falling real net prices (F26).
• Whether pre-K failure in Tennessee reflects quality or pre-K per se.

Unknown

• Whether administrative staff specifically outgrew instruction in higher ed. The coarse NCES data (F30) do not show it, and finer data were not verified.
• The effect of the 2025–26 federal downsizing on IES and NCES capacity (F34). Only NAEP's on-time 2025 release is established.
• US apprenticeship scale and P-TECH results: not verified this round.
• The effects of the FTCS (starts tax year 2027) and of RAP and the loan caps (effective July 1, 2026) (F20, F31).

Research agenda

1. Lottery or phase-in evaluation of every large tutoring rollout (built into P1).
2. Regression-discontinuity studies of 3rd-grade retention at cut scores, separating retention from instruction.
3. Linked attainment data for ESA/FTCS users vs matched public students, with tests and college outcomes. This would resolve the test-vs-attainment puzzle.
4. The Grad PLUS elimination as a natural experiment on graduate tuition (Finch's prediction).
5. Year 3+ follow-up of Florida's phone ban, with discipline disparities tracked.
6. A quality-controlled statewide pre-K RCT (Reyes's and Stroud's shared demand).

Consensus reforms (≥60% in all three blocs)

• P1: evidence-conditioned tutoring.
• P2: science-of-reading package without mandatory retention.
• P4: bell-to-bell phone rules with non-punitive enforcement.
• P7: higher-ed risk-sharing plus net-price transparency.
• P8: career academies and apprenticeship with RCTs.

Just missed or failed:

• P5 (testing of choice users) was one bloc short. The Right split 50/50 on it, the most interesting split in the thread.
• P3 (retention) and P6 (funding weights) failed on predictable lines.

Where Symposium 1 was wrong or too glib

I have not re-audited Symposium 1 line by line. But judged against the claims this round had to retract, the shallow versions of the education debate make five errors:

1. "Money doesn't matter" / "just spend more." Both are wrong. The defensible claim is that sustained money for poor kids helps on average, with unpredictable size.
2. "Tutoring will close the gap." Budget for 0.14 SD, not 0.4.
3. "Vouchers raise achievement" / "vouchers harm kids." On tests, modern programs have lowered scores. On attainment the evidence leans positive and is weaker. The honest summary is "mixed and program-dependent."
4. "Pre-K always pays." Tennessee's RCT forbids that sentence.
5. "College keeps getting more expensive." That is true of sticker prices, not of average net tuition at public four-years since 2012.

The most productive feature of this round was that all eight "Wrong" ratings came from professors whose direction was defensible and whose magnitudes were not. That is where policy mistakes are usually made.

— Wainwright

Independent Audit

Audit Panel: Prof. Laura Pennington (#97, operations research, C), lead, and Prof. Olivia Brandt (#57, statistics, C). Brandt sat on this thread's referee panel and wrote the Stage 4 pattern note. She recuses from every judgment below that concerns the Stage 4 ratings, the tally or the Brief's wording, and Pennington made those calls alone. All corrections below are appended. Nothing above is deleted or edited.

1. Source check (15 claims re-fetched 2026-09-24)

# · Claim · Cited figure · Verified figure · Status · URL
F1 · NAEP 2024 gr 4/8 · Reading −2/−2; math +2/flat; below Basic ~40/33/25/40%; only LA gr-4 reading and AL gr-4 math beat 2019 · Same; 4th-grade math below Basic is ~24% · Confirmed (rounding) · nagb.gov/news-and-events/news-releases/2025/…
F3 · Education Recovery Scorecard · 17/11/6%; 14.1% vs 3.9%; gap +11%; ESSER −10% of a grade · Same · Confirmed · cepr.harvard.edu/news/education-recovery-scorecard
F4 · Chronic absenteeism (AEI) · ~15 → 28.5 → 25.4 → 23.5% · Same · Confirmed · aei.org/…/lingering-absence-in-public-schools…
F5 · Jackson-Johnson-Persico · +0.27 yrs, +7.25% wages, −3.67 pp poverty; larger for low-income · Same · Confirmed · nber.org/papers/w20847
F8 · CALDER ESSER · +0.008 SD math per $1,000 (sig.); reading n.s.; $9–13k for full recovery · Same · Confirmed · caldercenter.org/publications/esser-and-student-achievement…
F11 · Kraft et al. tutoring · 265 RCTs; 0.42 pooled; ≥1,000 students (US, standardized) = 0.14; 400–999 = 0.21–0.25 · 0.42 pooled confirmed. 0.14 is the full-sample figure for ≥1,000; the US + standardized-test figure is 0.16. 400–999 is 0.25 (full sample) and 0.21 (US). "A third to a half" compares the US-standardized subset to the full sample, not scale to pilot. · Minor discrepancy · edworkingpapers.com/sites/default/files/ai24-1031.pdf
F12 · Mississippi NAEP · 219 vs 214; 239 vs 237; 253 vs 257; adjusted #1 in 3 cells, #4 in 8th reading · Same · Confirmed · mississippifirst.org/contextualizing-mississippis-2024-naep-scores/
F14 · Louisiana Scholarship · −0.4 SD math yr 1; negative in other subjects; low-quality school selection · Same · Confirmed · aeaweb.org/articles?id=10.1257/app.20160634
F16 · Ohio EdChoice attainment (Urban) · College 64 vs 48; BA 23 vs 15; matched design; modest spillovers · Same · Confirmed · urban.org/research/publication/effects-ohios-edchoice…
F19a · Arizona ESA (CSI) · 92,362; $1,036.9M vs $1,001.7M; $10,349 · Same. CSI also reports 57% of new enrollees came from public schools, which supports FC #22. · Confirmed · commonsenseinstituteus.org/…/esas-in-arizona-q3-2025-report/
F19b · EdChoice net cost · $178.3M / $118.0M; 1.1% / 0.7% · Same · Confirmed · edchoice.org/2025-did-arizonas-esa-expansion-blow-a-hole-in-the-budget/
F23 · EPI teacher pay penalty · 26.6% weekly (6.1% in 1996); 16.7% total comp; 9.9% benefits · Same. "Record" applies to the 26.6% wage penalty. · Confirmed · epi.org/publication/teacher-pay-in-2023/
F25 · Tennessee VPK RCT · 2,990 children; lower scores gr 3–6, worst in 6th; worse discipline, attendance, SpEd · Same; retention effect null · Confirmed (via EuropePMC; the PubMed link returned a CAPTCHA) · pubmed.ncbi.nlm.nih.gov/35007113/
F29 · Lucca-Nadauld-Shen · ~60¢ per $1 of subsidized-loan max; stronger at private/vocational · Same · Confirmed · newyorkfed.org/research/staff_reports/sr733.html
F30 · NCES Table 314.20 · Faculty 826,252 → 1,507,641; other staff 1,521,232 → 1,973,819 · Same (grad assistants 197,751 → 398,862, +101.7%) · Confirmed · nces.ed.gov/programs/digest/d23/tables/dt23_314.20.asp

Counts: Confirmed 14 · Minor 1 · Not supported 0 · Could not access 0.

2. Internal consistency

• Official tally verified. The "corrected count by line" is correct: S 30 · C 6 · U 4 · W 8 = 48, and every table row matches its list. The draft bullets above it (S 31 · C 7 · U 4 · W 8), which sum to 50, are superseded, and the post itself says so. For clarity, the split-rated rows #15 and #21 are counted once, as Supported, in the official tally.
• Vote math: all 24 yes-count/% pairs recomputed and correct, including 20/33 = 61% and 28/32 = 88% on P3 and 16/32 = 50% on P5. Consensus labels are correct: P1, P2, P4, P7 and P8 clear every bloc threshold, and P3, P5 and P6 fail. Marsh's "one of 16 Right yeses" matches.
• Bloc mislabel: the Stage 5 heading "*Center* steelmans Right on higher ed" names Vogt (#11), who is C-L, Left bloc. It should read "Left (C-L) steelmans Right on higher ed."
• FC #22 basis: the rating is Supported, but F19 as written in the Brief does not contain the "over half of new enrollees" figure. The correct basis is CSI Q3 2025 (57% switchers). The rating stands.
• F11 relabel: in the Brief, P1 and the Verdict, "0.14 SD at scale" is the all-programs figure. The US + standardized-test at-scale figure is 0.16. P1's budgeting assumption of 0.14 is conservative, so no vote is affected.
• Abernathy (Exchange E) says the 16.7% total-compensation penalty is "still a record." EPI's "record" language refers to the 26.6% wage penalty. The Verdict states this correctly.
• Wrong claims reappearing as established: none.
• Roster: all 16 named professors match the roster on number, field and lean, including #48 Kevin Marsh (education policy, R).

3. Balance

Wrong ratings split evenly between flanks (Right 4, Left 4). The Center received zero Wrong ratings, and one call looks lenient. Voss's "Mississippi's gains cannot be retention artifacts" was rated Contested, while Reyes's "Tennessee is a one-state fluke" was rated Wrong (as framed). Both are certainty overclaims that the cited evidence weakens or contradicts, and each speaker narrowed the claim on the record. By the thread's own standard, Voss's original wording would merit Unsupported or Wrong (as framed). This does not change any vote or the Verdict, which lists Mississippi as Contested. The Established list is well balanced: it records negative voucher test effects, which cut against the Right, alongside the negative Tennessee pre-K RCT, which cuts against the Left.

4. Grade: A−

Sourcing is very strong (14 of 15 claims exact), and the tally, the vote math and the consensus labels are all correct. Points come off for the Kraft label swap, the bloc mislabel in the Stage 5 heading, and a gentler standard for the Center's certainty overclaim.

— Pennington (lead); Brandt (recused on Stage 4 ratings, tally and Brief wording)

Verdict v2 (audit-corrected)

Prof. Adelaide Wainwright (#100, moderator). This post is appended to the record and replaces the Stage 8 Verdict wherever the two differ. Items marked "not verified this round" were re-checked on 2026-09-24.

1. Resolution of unverified items

Item · Where · Result · Figure and source
Wren: ESSER ≈ $190B, ≈ $3–4k per pupil, buying ~0.03 SD (FC #8, C because the totals were unverified) · Stage 3 · Now verified · ESSER I $13.2B, ESSER II $54.3B and ESSER III (ARP) $122B, a total of $190B (en.wikipedia.org/wiki/Elementary_and_Secondary_School_Emergency_Relief_Fund). Divided by ~49M public-school students, that is ≈ $3,900 per pupil, and 0.008 SD × 3.9 ≈ 0.03 SD. Caveat: this extrapolates CALDER's ESSER-III-based slope to all three rounds.
Wren: ESSER "largely" spent on HVAC, bonuses and reversed hiring (FC #9) · Stage 3 · Still unverifiable · The same source lists staffing, technology, mental health, construction and renovation, and stipends and bonuses among twelve allowable uses, but gives no spending shares.
Reyes: Perry/Abecedarian returned $7–$13 per $1 (FC #27, U) · Stage 2 · Partly verified / corrected · Perry: Heckman estimates $7–$12 per $1, and HighScope reports $7 (by age 27) and $13 (by age 40) per tax dollar (en.wikipedia.org/wiki/Perry_Preschool_Project, secondary). Abecedarian (ABC/CARE): the Heckman Equation reports a 13% annual return on investment (heckmanequation.org/resource/13-roi-toolbox/). That is a rate of return, not a dollars-per-dollar ratio.
Pierce-Marsh: IES evaluation capacity damaged (FC #41, U) · Stage 2 · Now verified · On Feb 10, 2025, ED cancelled IES contracts: DOGE says 89 worth $881M, and AERA/COPAFS say ~170. The cuts included the congressionally required NPSAS. ED said NAEP, IPEDS, the College Scorecard and College Navigator were not affected. insidehighered.com/news/faculty-issues/research/2025/02/12/900m-institute-education-sciences-contracts-axed
Boston lottery pre-K results (Contested #6, "not re-verified") · Brief · Now verified · Gray-Lobe, Pathak & Walters (NBER w28756): preschool enrollment "boosts college attendance, as well as SAT test-taking and high school graduation," reduces discipline including juvenile incarceration, and has "no detectable impact on state achievement test scores." Effects are larger for boys. nber.org/papers/w28756
Finch: falling net price financed by high-sticker discounting (FC #46) · Stage 3 · Still unverifiable · The NACUBO discounting study was blocked by robots.txt.
Finch: administrative staff outgrew faculty (FC #45) · Stage 2 · Still unverifiable · No occupation-level IPEDS data were fetched.
US Registered Apprenticeship counts; P-TECH RCT (F33) · Brief · Still unverifiable · The apprenticeship.gov and MDRC pages returned no figures.
ED program transfers; McMahon v. New York stay (F34) · Brief · Still unverifiable this pass · The fetch failed.
Pouch cost, testing cost, Mississippi coaching appropriation, NSC cohorts after 2017 · Stage 6 / Brief · Still unverifiable · Not load-bearing.
Audit: Kraft et al. at-scale figure · Audit · Corrected · 0.14 SD is the all-programs figure for ≥1,000 students. The US + standardized-test figure is 0.16 SD. The "a third to a half" comparison is between that subset and the full sample.
Audit: FC #22 basis · Audit · Corrected · The source is CSI Q3 2025: 57% of new ESA enrollees came from public schools.
Audit: Stage 5 heading · Audit · Corrected · It should read "Left (C-L) steelmans Right on higher ed." Vogt (#11) is C-L, in the Left bloc.

2. Rating normalization and tally

• #8: C → S, now verified, with the caveat noted.
• #46 was rated C only because it had no source, so C → U.
• #27: U → C. Perry's ratio is supported. The Abecedarian figure is an annual return, not a ratio.
• #41: U → S.
• #35, Voss's "cannot be retention artifacts": C → U on the audit's balance finding, because it is a certainty overclaim beyond the evidence. This matches how Reyes's "fluke" was treated.
• #5, #18 and #32 remain real disputes and stay C.

Tally: S 30 · C 6 · U 4 · W 8 → S 32 · C 4 · U 4 · W 8 (48 claims).

3. Corrected verdict

Established

• The decline is real and bottom-heavy. Every NAEP cell is below 2019. Grade 12 is at record lows. The 90th percentile held while lower percentiles fell. Recovery tracks district income by about 4x. Chronic absence is ~22–23.5%, about 1.5x its pre-pandemic level.
• Money matters on average, especially sustained money for poor children. Hanushek's median estimate is positive.
• ESSER had small positive effects at high cost. $190B, about $3,900 per pupil, bought ≈ 0.03 SD in math at CALDER's rate.
• Tutoring works, but much less at scale. Pooled RCTs show 0.42 SD. At-scale programs show 0.14 SD across all programs and 0.16 SD for US standardized tests. Budget for about 0.14–0.16.
• Modern statewide voucher programs have produced negative test-score effects.
• Tennessee's statewide pre-K RCT was negative through grade 6.
• Real net tuition at public four-years has fallen since 2012-13.
• Teachers face a record wage penalty of 26.6%. The total-compensation penalty is 16.7%, and the record language applies to the wage figure.
• New: The 2025 federal downsizing cut IES research contracts. About $881M were cancelled, including NPSAS. NAEP was protected and released on schedule.

Contested

• Voucher attainment effects vs test effects.
• Competition spillovers.
• Mississippi's gains: instruction vs retention vs composition. Voss's "cannot" is now rated U.
• Arizona's net ESA cost.
• The Bennett hypothesis.
• Pre-K: quality vs pre-K itself. Updated: Boston's lottery study, now verified, finds higher college attendance, SAT-taking and HS graduation, with no test-score effect, against Tennessee's negative RCT. That supports the "quality is the treatment" reading without settling it. Perry's $7–12 per $1 is sourced, but it comes from a tiny 1960s program.

Unknown

• Whether administrative staff outgrew instruction.
• The downstream effect of the IES cuts on RCT output and on data collections other than NAEP. That the cuts happened is now established.
• US apprenticeship scale and P-TECH results.
• The effects of FTCS, RAP and the loan caps.

Conclusions that change.

• (a) The Unknown item on federal capacity is partly resolved. The contract cancellations are established. Pierce-Marsh's claim moves from "prediction" to Supported, though the magnitude of its effect on research output is still unknown.
• (b) The pre-K Contested item gains verified positive evidence on attainment.
• (c) The tutoring budgeting figure is 0.14–0.16 SD, with no effect on P1.
• (d) Wren's ESSER arithmetic is sourced.

— Wainwright (#100)