Corrections log

The word “binding” fails the filing’s own test

Was Part 2 carried an unqualified “Binding constraint” stamp on workforce compensation, and the abstract opened on it: “the binding constraint on universal childcare in the United States is not appropriations but labor.” Halving turnover was the “arithmetically dominant lever,” with the whole sequencing recommendation — pay first, build second, extend third — resting on it.

Now Qualified, against the filing’s own definition of the word. Its research inquiry defines a constraint as binding only if relaxing it alone would materially raise achievable output and relaxing the others would not. The second clause fails. Re-running the filing’s own committed FTE model with nothing changed but the ratio parameters — set to what Florida’s statute already permits (Fla. Stat. § 402.305(4), roughly 7/17/25) — cuts required hiring from ~798,000 a year to ~501,000. Halving turnover, the lever the filing is built around, reaches 509,000 — and costs $45.9B a year, where the ratio change costs nothing and needs no federal vehicle. The model’s own infant ratio of 4.5 is tighter than the median state’s law and tighter than NAEYC’s recommendation while its docstring calls the values “licensing-typical,” and Part 5 praises Florida as a durable multi-decade success while modelling ratios 1.7–1.8× tighter than Florida requires. No scorecard dimension prices regulatory burden and no scored architecture touches ratios, so the alternative diagnosis was never tested, not ruled out. What survives: at the assumed ratios the required flow really is unprecedented, the five-country pattern holds, and low pay really does drive turnover — the steelman cannot claim regulation caused the low pay, because childcare’s wage relative to its own state’s median is uncorrelated with ratio stringency (r = −0.089). The stamp on the live page now reads “‘Binding’ is contested.”

“Three designs survive every ranking”

Was Part 6 published, as its headline, that three designs occupy the top ranks under all four weightings — a public option, a scaled Head-Start-style model, and K–12 extension — off a board described as “twelve candidate architectures … scored on fourteen anchored dimensions,” then re-scored blind.

Now The survival claim is qualified; one of the three is an artifact of the scoring. Three of the fourteen dimensions anchor their scale on the instrument category itself — “pure transfer” scores 1, “direct provision” scores 5 — so a cash or voucher design scores near the floor by definition, before any evidence is consulted. Drop those three and the top four changes under all four weightings: supply-first (a6) exits and caregiver-choice (a11) enters under three, and under equity-first a11 leads outright while the cash comparator — the design the anchor text cannot score above the floor — takes the fourth seat. This entry has itself been corrected. As published on 2026-08-10 it said “the defensible published claim is that a3, Head Start scaled to universal, leads under every weighting and every sensitivity tested — that finding is robust.” a3 leads no weighting, and did not on the day that was written. On the committed rankings — unchanged since the blind re-score of 2026-08-07 — a8, the public option, leads all four, with a3 second under three of them and third under the fourth. The steelman’s whole rank table had been computed against the pre-re-score board and was never re-run after that re-score moved 47 cells, two of them a3’s own (distributional 5→3, relief_workforce 4→3) — the cells that cost it the lead. Nothing new was discovered; a superseded result was carried forward for three days. The corrected statement is that a8 and a3 hold top-four places under every weighting in both specifications, and a3 leads none of them, while the rest of the top tier is weighting- and specification-dependent. A ranking script is now committed that reproduces the published rankings and refuses to compute any variant if they do not match. Also corrected, from Phase 1: the board is eleven architectures plus a cash comparator, not twelve architectures. The whitepaper had been counting the comparator as an architecture and the sources page had been counting twelve and then adding the comparator again; the score file has twelve rows — a1–a11 and a cash comparator. And the scales page’s claim that “every cell in the CSV carries an evidence tag resolvable in the rationale file” was corrected on its live mirror: of 168 cells, 9 are tagged.

What the Byrd rule actually did to the $15 minimum wage

Was “Reconciliation’s Byrd rule strips provisions whose budgetary effect is ‘merely incidental’ — the controlling precedent being the $15 minimum wage, struck from the 2021 relief act as a private-sector mandate. A childcare wage mandate dies identically; wage funding is money and survives.” The research record put the same rule more bluntly still: “anything that moves money survives; anything that mandates behavior dies.”

Now Rewritten on the procedural history, because almost every load-bearing word of that was wrong. The Byrd rule declares six categories of provision extraneous, § 644(b)(1)(A)–(F); the “merely incidental” test is one of them, and presenting it as the whole rule is what made the rest possible. Nothing was struck from the 2021 relief act — CRS RL30862 records “[none]” struck from P.L. 117-2. The provision was dropped before floor consideration after the Parliamentarian’s advice in February 2021; a separate amendment to add it back that March drew a point of order sustained on merely-incidental grounds, the waiver failing 42-58. The recorded ground was the budgetary effect being incidental to the non-budgetary one, not “a private-sector mandate,” which is not a category in § 644 at all — the sources page had it right and the error entered when the record was translated to the whitepaper. The Parliamentarian advises; the presiding officer rules; Senators remove. “Anything that moves money survives” is refuted by the filing’s own adjacent finding: § 644(b)(1)(E) makes a plainly money-moving provision extraneous if it increases outyear deficits — the very rule the filing invokes when it calls reconciliation-built sunsets “a design defect,” and CRS R48640 says the same in the other direction, that a provision with small budgetary and negligible policy effects can be permissible. “Dies identically” is now “faces the same objection,” because CRS records that every such determination is case-specific. The direction of the finding survives and is now properly hedged: wage funding is budgetary and clears the incidental test, though not automatically the outyear-deficit one. Separately flagged on the page: CBO publication 57630, which carries the $381.5B figure and the state-non-participation assumption, could not be obtained in primary form and stands as secondary-sourced.

The filing misnamed its own losers

Was Part 6 closed its scorecard summary on a rhetorical flourish: “The two most fashionable designs — unconditioned cash and employer cost-splitting — rank last.”

Now Half of that was right; the other half was contradicted by a file committed in the same repository, checkable in seconds. The committed rankings.txt never puts the cash comparator last under any weighting, in either scoring of the board. On the board as it stood at publication, cash finished 8th of 12 under the equal weighting (2.79) and no lower than 10th under any of the four; on the re-scored board now committed it finishes 7th of 12 under the equal weighting (3.14), again no lower than 10th anywhere. The bottom two are the same two rows under all four weightings and in both scorings: a5, a pure demand-side allowance (11th) and a10, Tri-Share federalized employer cost-splitting (12th). Employer cost-splitting therefore does rank last — a10 is that design — and unconditioned cash, the other name in the sentence, is mid-table. Nothing was wrong with the scoring; what was wrong was the sentence written about it, shaped for the point it wanted to make rather than read off the file underneath it. Part 6 now reads “a pure demand-side allowance and federalized employer cost-splitting — rank last; unconditioned cash is mid-table.”

“No mechanism grants an outside party standing access”

Was The abstract and Part 3 both stated, with no country attached, that no mechanism grants an outside party standing access to a lab’s weights, training data or compute logs, and that evaluators’ access is “negotiated per engagement… never standing rights.” The scorecard’s axis-C note said no counter-example “was found.”

Now The unqualified claim is withdrawn and re-scoped to the United States. EU AI Act Article 92 lets the AI Office conduct evaluations of a general-purpose AI model, lets the Commission “appoint independent experts to carry out evaluations on its behalf,” and lets it “request access to the general-purpose AI model concerned through APIs or further appropriate technical means and tools, including source code” — with Article 101(1)(d) setting the price of refusal at 3% of worldwide turnover or €15,000,000, whichever is higher. Article 91 and Annex XI reach the other two items: training-data provenance, and “the computational resources used to train the model (e.g. number of floating point operations).” It binds American frontier labs because they place general-purpose models on the EU market. Chapter V applied from 2 August 2025 and Articles 88–94 and 101 from 2 August 2026 — two days before this filing published. The Digital Omnibus on AI amends none of them. What is not established, and is said wherever the source is used: no exercise of the power is documented anywhere this pass reached, and Article 92(3) names source code, not model weights. The American finding stands and is reinforced: California’s SB 53 and New York’s RAISE Act were both read in full at primary tier, and neither contains the word “audit” or grants access to anything — and RAISE, on the replacement chapter that actually governed at filing, does not take effect until 1 January 2027.

“Every licensure proposal currently pending relies on self-certification”

Was Part 9 rested the frontier-licensure placement on what is “currently pending,” and the research record described New York’s RAISE Act as a pending bill that “nominally requires third-party audits” whose durability through amendments was “uncertain.”

Now Corrected on the enacted text, in the direction that makes this filing’s own safety finding stronger. The RAISE Act is not pending: S6953-B was signed 19 December 2025 as Chapter 699 of the Laws of 2025. Searched in full, General Business Law article 44-B contains no third-party audit provision; the word “audit” does not appear in it. The durability question the record hedged had already resolved, and it resolved against the audit. The rest of this correction — its effective date, its penalty figures, and its claim that the act compels on-demand Attorney-General access to the unredacted safety protocol — was drawn from a chapter that had already been repealed, and is superseded by the follow-up below.

Follow-up · 2026-08-10 This correction was itself corrected, on the same day, because it read a repealed chapter as the operative law. The steelman fetched Chapter 699 and never asked whether Chapter 699 was still Article 44-B. It was not: A9449 / S8828 repealed and replaced General Business Law article 44-B in full, signed 27 March 2026 as Chapter 96 of the Laws of 2026, four months before this filing published, realigning the article to the structure of California’s SB 53. On the law that actually governed at filing: the effective date is 1 January 2027, so RAISE was not in force on 4 August 2026 and California’s SB 53 was the only American frontier statute then operative; the penalties are $1M for a first violation and $3M per subsequent violation, not $10M and $30M; and the Attorney General’s on-demand access to the unredacted safety protocol — the strongest thing the correction above claimed for the act — has no counterpart in the replacement at all. Where the repealed text gave a regulator a right to obtain the unredacted document, the current §1421(5) gives the developer a right to redact, requiring only that it “describe the character and justification of such redaction” and keep the unredacted material itself for five years. Incident reports now go to an office inside the Department of Financial Services rather than to the Attorney General. What survives both versions, checked in full against the current text rather than carried over: the word “audit” appears nowhere in article 44-B — nor does “auditor,” nor “independent” — and the 72-hour safety-incident reporting duty stands, joined by a new 24-hour duty where an incident poses an imminent risk of death or serious physical injury. The only third-party language in the statute requires developers to disclose whether outside evaluators were used, never to obtain one. The filing’s underlying finding is unchanged and reinforced: no American mechanism grants an outside party standing access to a lab’s weights, training data or compute logs — true a fortiori of a statute that grants access to no one and does not operate until 2027. Only the supporting figures moved, and no scorecard cell moved.

Three scorecard results that each turn on one cell

Was The abstract and Part 9 published “state-law primacy and disclosure mandates co-leading,” “frontier-lab licensure sharing the floor with the voluntary status quo,” and “the comprehensive federal statute mid-pack as an all-neutral row” — as results of a nine-architecture, six-weighting board corrected twice.

Now Both readings published, and the fragility stated. Move frontier licensure’s concentration cell from 2 to 3 — the correction the blind re-score already applied to the identical unmeasured “structurally favors incumbents” claim one row up, and never applied here — and “shares the floor on four of the six weightings” becomes true on none. Move disclosure’s harm cell from 4 to 3, the same correction the red team applied when it stopped a row being credited for the size of the problem it would address, and the co-lead ends: state-law primacy leads alone on all six. Move the comprehensive statute’s catastrophic-risk cell to 4 on the EU evaluation power above, and the row published as mid-pack co-leads four weightings and takes the catastrophic-risk weighting outright. One result held under every variant: the voluntary status quo still scores worse than doing nothing, on all six weightings, even at the corrected reading of the cell it rests on. Nothing was re-scored on the steelman’s own authority — a steelman pass is not a third blind scorer — and a ranking script is now committed that reproduces all fifty-four published values before computing any variant. A structurally blinded re-score of the four contested cells is owed. Also flagged: an architecture with no real implemented instance cannot exceed 3 on any axis under this rubric, so its ceiling is the mid-pack and it can only move down — six of nine rows are in that position.

“The one thing no bill proposes”

Was Part 10’s plain-talk box said the approach this filing ranks first — make exchanges keep customer assets separate and prove it to auditors — “is the one thing no bill proposes.”

Now Half of it is in the bill, in terms. CLARITY Act §404, writing a new CEA §5i(d)(2), requires a digital commodity exchange to treat all customer money, assets and property “as belonging to the customer,” “separately accounted for,” with commingling “prohibited”; §406 puts the identical duty on digital commodity brokers and dealers — which is what Celsius and Voyager actually were; and §402 amends 7 U.S.C. 6d itself, the 1936 futures statute the language is transplanted from near-verbatim. The same bill makes holding segregated customer funds the trigger for compulsory membership in a registered futures association and directs that the Commission “shall require” that body to write “such rules as may be necessary to further compliance” with the segregation duty. Phase 1 had fetched this bill and read its custody title for the audit question only; the segregation provisions sit in different sections and were never searched for, and the bill was never committed to the source atlas, so no later pass could read it from disk. What survives is narrower and sharper: the bill sets no cadence, no proof-of-reserves standard and no method for confirming customer assets are there — searched literally, “proof of reserves” appears zero times, “reconcile” in any form zero times, “concentration” zero times. It creates the duty and delegates the verification. Crediting the segregation duty moves no ranking on the scorecard.

“No as-implemented exemplar anywhere on earth”

Was This filing’s most-repeated disclosure — in the honesty box, in Part 4, on the sources page and in the scorecard — was that the audit-with-teeth half of its leading architecture, the half it calls load-bearing, has no as-implemented exemplar anywhere on earth.

Now Withdrawn, and the exemplar makes this filing’s own argument stronger than the version it published. American law has run exactly this experiment. 7 U.S.C. §6d(a)(2) has required futures brokers to treat customer money “as belonging to such customer,” “separately accounted for and … not … commingled with the funds of such commission merchant,” since 1936 — and Peregrine Financial Group defeated it by lying about the balance, telling an NFA examination in July 2012 that it “held in excess of $220 million of customer funds when in fact it held approximately $5.1 million.” The verification half was adopted after that and after MF Global’s reported $900 million shortfall: 78 FR 68506 (Nov. 14, 2013, effective 13 January 2014) requires the depository to give the regulator “direct, read-only electronic access to transaction and account balance information for futures customer accounts,” in an acknowledgment letter the depository must file “without further notice to or consent from the futures commission merchant,” with 17 CFR §1.32 requiring the segregated-account computation. That is this filing’s own thesis — the intermediary’s report of its customer balances is not trusted — implemented for twelve years, in the agency and the self-regulatory body the CLARITY Act would put in charge of crypto exchanges. The claim came from a red-team attack that correctly said “nowhere in the record,” a statement about this project’s own research, and was published as a statement about the world; nobody tested the negative against primary law. The limit, stated so nobody over-reads it: no study measuring whether futures-broker customer-fund shortfalls fell after 2014 was located, so this is design evidence, not demonstrated effect — and no scorecard cell was raised on it.

“Leads every weighting under every sensitivity reading”

Was Part 10, the scorecard’s own finding 1, the sequencing memo and two cards on the sources page all said the architecture this filing ranks first — custody with audit teeth, #4 — leads every weighting under every sensitivity reading, and that it is “strictly first under all four A×E combinations.”

Now Contradicted by our own printed table, in the same document, for four days. Under the enforceability-first weighting (W5) the do-nothing comparator leads outright, 3.25 to #4’s 3.15 — a number the scorecard has carried in its rankings table, and stated in its own “named artifact” prose, since 6 August 2026, in the same file whose finding 1 and sensitivity list claimed the opposite. Nothing new was discovered: the claim was true of an earlier version of the board and was never recomputed after adjudication moved the do-nothing row’s enforceability cell (8E) from 3 to 4. Phase 1 had re-derived the row sums and recorded them as confirmed, correctly — the error was in the claims about the table, not in the table. Two further defects sat alongside it: the rankings table was headed “Primary reading (A=3, broad E)” while printing the A=4 matrix, and the sensitivity list offered an “upper reading” that v4 had already made primary while offering no lower sensitivity at all for 4A — the single cell the entire lead rests on. All six weightings have been recomputed from the committed v4 cells across five readings, and the “v4 as published” row reproduces the committed table exactly, which is the check that the rest is trustworthy. The corrected statement is that #4 leads five of the six weightings and loses the sixth to doing nothing — in the primary reading and under every sensitivity now published. And the lead is one cell wide. Return 4A to 3 — the red team’s own reading, held by one of the two blind re-scores — and #4 leads nothing outright: it ties #3 across W0–W4 and still loses W5. What does not move: outright restriction (#7) finishes last under every weighting in every reading, and the tie between the marquee bills and doing nothing is untouched.

“Medicare mostly doesn’t pay for this”

Was The stamped verdict on every screen — rail, masthead dossier, homepage cards — read “Medicare mostly doesn’t pay for this,” treating the LTSS construct (Medicaid 61% of $415B) as the whole story of who pays when an older person needs care at home or in a nursing facility.

Now Narrowed. Medicare does not fund custodial long-term care — 42 U.S.C. §1395y(a)(9) bars it — and that half survives. What does not: under CMS’s National Health Expenditure Accounts for 2024, Medicare is the largest single payer of home health care in the country (32.9%), ahead of Medicaid, and it already runs nationwide caregiver-respite and dementia-care management through GUIDE and Medicare Advantage supplemental benefits. The rail and the public pages now read: “Medicare doesn’t pay for custodial care. It pays for plenty of the rest.” Two legitimate constructs; the filing had published only one.

“The best-evidenced fix isn’t the one anyone’s proposing”

Was Part 9’s heading framed Cash & Counseling / self-direction as the board’s best-evidenced design and as something nobody was proposing — the contrast that made the scorecard’s lead feel like an original finding.

Now Withdrawn as false. Congress enacted that design as a permanent state-plan option in 2005 (42 U.S.C. §1396n(j)) and again in the ACA as Community First Choice (§1396n(k)), which pays states a six-percentage-point federal match bonus to adopt it. Self-direction now runs in all fifty states and DC; about 1.5 million people use it. The research file already said so; the page said the opposite. The one study that evaluated the statutory version found nothing — the same Ne’eman 2026 paper cited for the board’s strongest positive cell reports no significant workforce effect for Community First Choice beside BIP’s +13.24%, and only the BIP arm had been published. The live heading now reads that the best-evidenced fix is already law in every state, and that nobody has measured whether scaling it worked.

“Private LTC insurance reform ranks last under all three weightings, with no exceptions”

Was The filing’s most confident negative finding: private long-term-care insurance reform ranked last under every weighting, “the clearest, most stable negative finding in the whole scorecard.”

Now Withdrawn. The structurally blinded re-score found that four of that row’s five cells rested on no evidence either way — they had been marked down because the old scale treated “no evidence found” as a low score while a high score required evidence. Corrected, thirty of forty cells sit at dead neutral; the row is no longer a stable last-place finding. Federal LTC social insurance was moved into the last-place seat on the corrected board, and even that replacement headline is published with a sensitivity: one below-neutral workforce cell is an ageing-country shortfall the United States also has, and scoring it at neutral turns “ranks last alone” into “ties for last.”

Long-term care need rate

Was Protocol prior cited a superseded figure — roughly 70% of Americans turning 65 will need long-term care.

Now Research corrected the load-bearing figure to 56%. The wrong number stays visible in the anchor table next to what turned out to be true.

Architecture that “survives all rankings”

Was First scorecard draft claimed one architecture won under every weighting.

Now Red-team pass found the score behind that claim had never been verified. The claim was withdrawn rather than defended — stamped as corrected via red team, not asserted.

How many justices backed the platform-speech holding

Was Moody v. NetChoice held “with a genuine six-justice majority” that a platform’s curation of its own feed is protected editorial speech — “not a minor aside, a real holding.”

Now Five. The reasoning we relied on is in Part III–B of the opinion, which Justice Jackson did not join — she took Parts I, II and III–A only, and wrote that she “would not go on to treat either like an as-applied challenge and preview our potential ruling on the merits.” Three more justices called the whole discussion “nonbinding dicta.” The Court’s actual holding was to vacate and remand. We had the vote breakdown right in our own record and drew the wrong inference from it.

Where CPB’s insulation actually came from

Was CPB’s two-year advance, formula-based funding was “a firewall built into the Public Broadcasting Act of 1967 and never removed.”

Now Not in the 1967 Act. The funding mechanism was built across the Public Broadcasting Financing Act of 1975 and the Public Telecommunications Financing Act of 1978, and the forward funding was settled practice rather than statutory command — Congress appropriated CPB’s FY2026 money, $535 million, back in 2024. This makes the collapse to a simple-majority rescission less surprising than we said, not more.

Canada’s local-versus-national engagement split

Was “Local outlets lost 85% of their Facebook engagement at the one-year mark versus 64% for national outlets,” sourced to the Media Ecosystem Observatory’s one-year report.

Now Withdrawn. That report contains no national-outlet figure and no local-versus-national comparison at all; its 85% is an all-outlets number. The asymmetry finding survives on a figure that is in the source and is local-specific: 30% of the 713 Canadian local outlets previously active on social media went dormant, and 212 of the 217 outlets that went dark — 98% — were local.

The citation behind Australia’s “90% to three incumbents”

Was “The one peer-reviewed distributional estimate available,” cited as Brevini & Ward, Media International Australia (2023).

Now No such paper exists. The real work is Brevini, sole-authored, in Javnost — The Public 30(2) — a different journal and no co-author. It is paywalled, and our own deviations log records that we never read it. The 90% figure now carries that caveat on the page instead of a peer-review badge it hadn’t earned.

The news-desert county counts

Was 213 counties with zero local outlets; 1,524 with exactly one.

Now 212 and 1,525, per Medill’s own report and dataset; 213 appears only in its landing-page essay. The two pairs sum to the same 1,737 counties, so the ~55% share — which is our arithmetic, not a Medill-stated figure — is unchanged.

“The one architecture the record actively refutes”

Was Part 9 said the platform bargaining code “is the one architecture the record actively refutes on four of five axes, not merely ranks last,” with each of its 1-scores resting on peer-reviewed or platform-measured evidence. Row 2 sat alone at the bottom at a neutral mean of 1.40.

Now Withdrawn. Three of those four bottom scores rested on the ~90%-to-three-incumbents figure — cited to a paper that does not exist, drawn from deal values the Australian Treasury’s own statutory review says it could not obtain. That same review concludes “it is reasonable to conclude that the Code has been a success to date,” documents agreements covering small and regional publishers, and records the ABC using the proceeds for regional reporting posts in places that had none. Canada put a mandatory per-journalist distribution formula with hard incumbent caps into law in 2023; Australia doubled its small/regional loading on this filing’s own filing date. Re-scored on what survives, the row moves to about 2.4–2.8 — mid-board, not refuted. Nothing above it changes rank. Meta’s exit still stands as the strongest surviving objection.

The legal cap behind the filing’s own top recommendation

Was Contingency management was “capped … by a federal anti-kickback rule one agency could lift tomorrow,” held “for years at a third of its trial-effective incentive level by a federal anti-kickback policy one office can change on its own,” with a federal rulemaking “already underway.” That mechanism is what made CM the top-ranked architecture in Part 6 and the filing’s first move in Part 7.

Now Withdrawn as described. OIG’s own 2020 final rule (85 FR 77684, at 77791–92) says it directly: “There is no OIG-imposed $75 limitation on contingency management program incentives.” The $75 is nominal-value guidance under the Beneficiary Inducements CMP — a different authority, in-kind only, and one OIG states “applies only with respect to the Beneficiary Inducements CMP and not to the Federal anti-kickback statute.” The $75 that actually bound programs was SAMHSA’s own grant condition, which SAMHSA lifted to $750 in January 2025. The real regulatory ceiling is the patient-engagement safe harbor at 42 CFR 1001.952(hh)(5) — a $500 base, CPI-indexed to $623 for 2026, in-kind only, available only inside a qualifying value-based enterprise. And OIG cannot repeal the statute by rulemaking; it can only add a safe harbor, and only “in consultation with the Attorney General” (42 U.S.C. §1320a-7d(a)(1)(B)).

Now The rulemaking is not underway either. The November 2024 document the filing cited is 89 FR 93545 — OIG’s statutorily required annual solicitation for safe-harbor proposals, published in 2022, 2023, 2024 and again in December 2025, which does not mention contingency management anywhere. The actual CM safe-harbor rulemaking is RIN 0936-AA13, sitting in Long-Term Actions on the Unified Agenda with an NPRM target of July 2027. Three scorecard cells that rested on this mechanism (A3 D2, A3 D8, A1 D2) are now contradicted by their own basis text and are recorded as owed a fresh blinded re-score rather than patched here.

The June 2025 “deaths rising again” reversal

Was Part 3 said the widely reported June 2025 reversal “turned out to be a statistical artifact in CDC’s own forecasting model — corrected months later, after the story had already spread.”

Now Withdrawn by the Phase 1 pass, on two grounds: that the cited article’s title — “The 2025 Drug Overdose Spike That Wasn’t: Neither Politics nor Data Errors Explain the Anomaly” — asserted the opposite, and that its body was paywalled and could not be read.

Follow-up · 2026-08-10 The withdrawal was itself wrong, and is reversed the same day. An independent adversarial audit fetched the source instead of its title. Both premises were false: the article is open access at PMC13066679 and says, verbatim, “Instead, the anomaly was a model artifact”; that CDC’s “second revision released in August 2025” clarified “that the January 2025 ‘spike’ was an artifact”; and that “the anomaly resulted from applying growth-era algorithms to a period of decline.” The title rules out political manipulation and coding or reporting error — which leaves the forecasting-model artifact, which is what the filing said. The claim is restored and confirmed at primary-source tier on both public pages and in the anchor table; four ledger verdicts moved back (A53, B11, C14 corrected → confirmed; A54 unverifiable → confirmed). What survives from the withdrawal is the half that was right: the citation is Post et al., Am J Public Health 2026;116(5):591–593, not 2025. The failure mode has a name now — a verdict rendered on citation metadata plus an untested paywall assumption. A correction is a claim like any other and carries the same primary-source burden: withdrawing a true claim is as much an error as asserting a false one.

The binding constraint is law — wrong lever, wrong venue, wrong clock

Was The abstract opened on it: “The binding constraint on American drug policy is not evidence — it’s law.” Contingency management was boxed in by federal fraud-and-abuse rules; the sequencing recommendation put a federal safe-harbor rulemaking first; the scorecard treated CM as the architecture the law was holding back.

Now The steelman won on the part that mattered most. The filing’s own evidence already said otherwise: the buprenorphine waiver repeal added prescribers without adding patients; mobile methadone units ran into staffing and community resistance rather than statute; and the VA delivers CM to 1.2% of the patients it diagnoses with stimulant use disorder despite facing no OIG safe-harbor exposure — it has its own incentive cap under IRS reporting rules, so caps on incentive size bind generally rather than “remove the law and uptake still stays low.” Five states have §1115 Medicaid CM approvals; California has been delivering since March 2023; CMS-approved incentive maxima run up to $1,092 over 24 weeks. The evidenced fastest route is a state §1115 amendment, not the federal safe-harbor NPRM targeted for July 2027. What did not fall: Part 4’s enforcement finding, Part 3’s multi-cause overdose decline, and the case for CM as a treatment — the first real-world mortality evidence (2025) is favourable. The correction is not “CM was a bad idea.” It is that the filing named the wrong lever, the wrong venue and the wrong clock, then built sequencing on top of all three.

Corrections that stopped at the research file

Was The research record corrected the CM legal status on 2026-08-06 — deviation #20, logged in ws14-cm-legal-status.md, which says in terms that it “supersedes any statement that CM has a settled nationwide safe-harbor expansion or an OIG $75 ‘cap’.” The same day, ws12-switzerland-root-trace.md closed the Switzerland gap and closed it against the filing. Neither edit reached the public pages; ws12’s own note deferred the site edit pending PR #39, which closed without the edit landing.

Now Both carried to the page, four days late. The live whitepaper ran a superseded legal claim as its headline recommendation for four days, and its honesty box told readers Switzerland was “still open” after the record had already traced it and found it refutes the case’s use as a positive counterfactual — a four-pillar framework that retains enforcement, whose exceptional intervention is heroin-assisted treatment for roughly 1,700 treatment-refractory patients, not decriminalization. Also corrected in the same pass: methadone’s OTP monopoly is not “written directly into statute” (21 U.S.C. §823(h) requires a separate annual registration and is facility-neutral; the clinic-only structure is 42 CFR Part 8 and DEA registration standards built under it), Oregon’s null-result count falls from “three of four peer-reviewed studies” to two — the third, an April 2026 changepoint study, could not be located in any index — and a quotation attributed to Spencer 2023 is not in Spencer 2023 — the abstract says “in a context with relatively poor access to drug treatment services,” not “not accompanied by substantial public health investments.” A record correction that stops at the research file has not been made.

“Strongly cautions the bare voucher entitlement”

Was The scorecard scored the voucher entitlement at the floor of the mobility axis (“documented to fail precisely in opportunity areas”) and as unevidenced on stability, and Part 9 published it as one of two instruments the record “strongly cautions.” Sequencing put demand-side assistance last.

Now Narrowed to the voucher’s supply effect; withdrawn as a verdict on the instrument. The two largest randomized evaluations of the US housing voucher appear nowhere in this filing’s record. HUD’s own Family Options Study — 2,282 families across 12 cities, randomly assigned, with the treatment arm receiving an ordinary housing voucher and no other services — cut homelessness from 38% to 17% on a pre-registered confirmatory outcome, and “more than halved most forms of residential instability.” The Creating Moves to Opportunity trial (American Economic Review 114(5), 2024) raised moves to high-opportunity neighbourhoods from 15.4% to 53.2% at $2,670 a household, using a bundle that includes the landlord engagement this filing’s own denial finding calls for — while a straight increase in what the voucher pays in expensive areas moved almost nobody. The caution stands only for what it was always about: a voucher builds no housing. “Demand-side last” keeps its preconditions and loses its rationale.

“Leads every weighting tested”

Was Part 9 stated flat that envelope-prescriptive legal reform “now leads every weighting tested,” with no sensitivity — even though the scorecard file promised one and did not print it.

Now Qualified, and the arithmetic committed. Twenty-three of the board’s thirty-two cells are neutral, and the leader holds the only cell above neutral anywhere on it — so its lead follows from the shape of the matrix rather than from running the weightings, and is worth a quarter of one point under the neutral weighting. Under the red team’s own stricter reading of that single cell — the cell the filing’s red team argued down to 3 and its blind second scorer marked 5 — it leads none of the 1,771 weightings swept, and ties the land-value tax, the row this filing itself calls an ignorance artifact. A ranking script is now committed so the claim is reproducible; Phase 1 had recorded it as unverifiable because none was.

The scorecard has no instrument for one of its own four objectives

Was Nine architectures scored on four objectives, with three adversarial checks reported (red team, blind re-score, Phase 1 verification).

Now Disclosed. The nine-architecture list was written before the first workstream ran and no workstream ever added to it. The board therefore carries no tenant-protection or anti-displacement instrument, although stability is one of its four scored objectives and the filing verified both the doctrinal ground (rent regulation survives facial takings challenge — Block, Pennell, Yee, Lingle, read as opinions) and the political ground (two enacted in the window). It carries no preservation instrument either, although Part 6 is about a constraint that destroys existing affordable units. All three prior checks audited cells; none audited rows.

“More than half of all multifamily operating-expense inflation is insurance”

Was Part 6 led with “more than half of all multifamily operating-expense inflation since 2020 is property insurance alone,” attributed to the Minneapolis Fed.

Now Withdrawn. Fetching the article found the figure is a quotation from a single anonymous respondent to a survey of multifamily owners, about their own portfolio: “We estimate that over 50 percent of our overall operating expense inflation since 2020 can be explained by property insurance premium increases.” It is not a Fed measurement and not a market aggregate. What survives, verified in the same article: premiums roughly doubled 2021→2024, up ~45% in 2023–24 alone, more than six times CPI.

The “46×” design-detail multiplier

Was “A design detail worth a 46-fold difference in real production” — in the page title, the social card, the abstract and the Part 2 stat strip, labelled “prescriptive vs. discretionary uptake.”

Now Withdrawn. The multiplier is asserted in six places in the record and derived in none. No pair of quantities in the supporting table has a ratio near 46; the only arithmetic that produces it is 25,000 ÷ 540, which is California’s ADU program’s own 2016→2022 growth — not a comparison between reform designs. The underlying finding is unaffected and now leads with the committed admin-data figure instead: Austin’s three-to-four-unit permit class ran 51 → 8 → 18 units across the reform that legalized it (2023 → 2024 → 2025).

Follow-up · 2026-08-10 An independent audit of this correction found the replacement itself incomplete: the stat strip published only the 51 → 8 trough, cutting the 2025 recovery to 18, and a superseded tile (“~15 Austin HOME 3-unit permits in 2 years”) was left standing beside it, where the two figures read as a contradiction because one counts permits and the other counts housing units. The series is now published in full with the permit-count figure moved into the prose that distinguishes it, and both published limits travel with the stat: the three-and-four-family class attribution is confirmed by a unit-count query against Austin’s permit data rather than against Austin’s written classification guidance, and the same series carries a classification break (two-family 156 → 532 while secondary-apartment collapses 248 → 10).

“1.4M → 1.1M public housing, peak → today”

Was The Part 7 stat strip put today’s public-housing stock at 1.1 million units.

Now 878,000. The 1.1M traces to a CRS report last updated in 2014. HUD’s own Faircloth List — the same file, the same sheet from which this filing computed its 278,345-unit headline — gives 878,000 standing units, and the filing’s own arithmetic requires it: 1,156,358 current limit − 878,000 standing = the 278,345 of headroom. At 1.1M the headroom would have been about 56,000.

“A net-increase freeze at 1999 levels”

Was The Faircloth limit described as “actually a net-increase freeze at 1999 levels.”

Now Read against 42 U.S.C. §1437g(g)(3) itself, it is neither a cap nor a freeze: it is a condition on the use of Capital and Operating Fund money, and subparagraph (B) expressly permits a housing authority to build above the line anyway, without extra formula funding. HUD’s own file says as much in its title — “Maximum Number of Units Eligible for Capital Funding and Operating Subsidy.” This cuts in the filing’s favour on the conclusion and against it on the mechanism.

The sources page never received the 2026-08-06 re-score

Was The public research record still led with the withdrawn v2 finding — “a stable pair converging on a trio,” the Faircloth channel leading stability — and its status read “Draft pending second scorer.” The re-score commit updated the whitepaper and not this page.

Now Updated to the reconciled matrix: one evidenced leader across all six weightings, the Faircloth stability lead withdrawn, status “Twice checked.” Also corrected on the same page: deviations count 65 → 69, workstream count 11 → 13, method import M1–M9 → M1–M8, and anchor 2’s single-root caveat, which the record had already closed.

“3.8–31× required workforce growth”

Was Presented as the range across any defensible shortage band.

Now 2.6–63×. The published pair is one row of the committed model — the ten-year ramp. The model’s own least-demanding case is 2.6× and its most demanding is 62.7×. The qualitative finding is unchanged and stronger: no combination of band and ramp falls within trend.