GBMT-8 (Drugs) — Phase 2 Steelman Log

Protocol: method/verification-protocol.md, Phase 2 (S1–S5). Filing under test: Whitepaper No. 8, filed 2026-08-03, as corrected by Phase 1 (2026-08-10, 393 claims, 57 moved). Date: 2026-08-10.

Independence (S1). Run in a session that did not author the filing, did not run its red team or its blind re-score, and did not run its Phase 1 fact-check. The Phase 1 log was read as the corrected state of the record, not as a map of where to attack; the tilt below was derived from the filing's own hypothesis set, scorecard and conclusions before any steelman was built. Where a steelman lands on ground Phase 1 already disturbed, that is said plainly rather than re-presented as a new discovery.

S2 — The tilt, derived from the filing's own text

Which direction it leans. The filing's thesis is stated in its first sentence: "The binding constraint on American drug policy is not evidence — it's law." Everything follows from that. If law binds, the highest-value moves are legal changes; if legal changes are cheap when they are regulatory, the best move is a federal rule; and the best federal rule is the one attached to the best-evidenced treatment. That chain produces the filing's recommendation without any further input.

The protocol pre-committed to it. research-inquiry.md §2 calls H4.3 — "the interventions with the strongest evidence … are the ones most constrained by law" — "the filing's likely headline shape," before fieldwork. H6.1 predicts the binding rules are "mostly federal." H6.2 predicts "a material fraction of the treatment agenda is executable by rulemaking alone." H10.1 predicts "the strongest-evidence architectures are legal-constraint-bound, not funding-bound." Four hypotheses, one direction, and the filing reports all four as supported.

Which architecture won. A3, contingency-management safe-harbor expansion — the narrowest, cheapest, most purely legal instrument on the board. It scores 3 on evidence, 3 on legal executability, 3 on cost and 3 on speed.

Whose framing it adopted. The federal regulatory analyst's. Policy is a stack of federal legal objects — a statute, a safe harbor, a scheduling order, an agenda entry — and the question is which object can be moved by which office. Delivery systems, payers, clinics, staff and patients appear in the record (§4 documents them well) but not in the thesis.

What it argued against, scored down, or never entertained. Three things.

1. The competing diagnosis. "Delivery capacity and payment bind, not law" is never scored as a position, even though the filing's own §4 contains the cleanest natural experiment against its thesis — the X-waiver repeal, which the filing describes accurately in Part 2 and then does not let generalise.
2. The architecture its own masthead vindicates. The filing leads with "the legal drugs kill more" and then ranks A9 — the only architecture aimed at the legal drugs — last, on the ground that D4's anchor says "overdose deaths specifically." The re-score flagged this as "a scoring-frame limitation" and left the rank in place.
3. Whether invisibility is a virtue. Part 7's sequencing logic is that a politically invisible rule "can bank results before any backlash coalition has time to organize." The opposite reading — that invisible rules have no constituency, so nothing forces them onto a crowded agenda — is nowhere in the record. A3 scores 3 on political durability on the strength of the first reading.

Therefore three steelman targets:

Note what is not a target. Enforcement is the direction M3 required the filing to steelman, and it did: A10 was raised to Σ18 on reconciliation, higher than nine of eleven rows. Assuming the tilt from the subject matter — "a drug-policy filing must be soft on enforcement" — would have produced a strawman. S2 says derive the tilt from the text. The text leans regulatory, not permissive.

S3 — Steelman A: the binding constraint is delivery and payment, not law

Built to persuade a reader who currently holds the filing's conclusion. Every figure below was read inside a source fetched this pass; nothing is taken from a search summary.

The claim. Removing a federal legal barrier to a drug treatment reliably produces more permission and unreliably produces more treatment. Where the barrier has actually been removed — three times in four years, at national scale — the outcome moved little or not at all, because what binds is the staffed, paid, supervised apparatus that delivers the service. The filing recommends removing a fourth barrier and scores the result as fast, cheap and high-impact. Its own evidence base says that is the least likely of the three to be true.

A-1. The buprenorphine waiver: the experiment already ran, and the filing reports it

The Consolidated Appropriations Act, 2023 eliminated the DEA waiver requirement on 12 January 2023 — the exact intervention class the filing recommends, executed at national scope. Chua, Bohnert and Nguyen (N Engl J Med 2024;390(16):1530–1532) measured it against IQVIA data covering 92% of US retail dispensing. Read from the source this pass (the figures Phase 1 could not reach; NIH author manuscript, PMC11103581):

"In January 2023, there was a level increase of 1938 prescribers (95% confidence interval [CI], 990 to 2887) and a slope increase of 595 prescribers per month … In December 2023, there were 53,635 buprenorphine prescribers."

"In contrast, between January and December 2022, the monthly number of patients to whom buprenorphine was dispensed increased only slightly, from 810,911 to 831,656. In January 2023, there was no substantial level change in the number of patients with prescriptions (2683; 95% CI, −6751 to 12,117) or slope change (−857 patients per month; 95% CI, −2943 to 1229)."

The authors' own conclusion:

"These findings suggest that the policy may have reduced barriers to prescribing but was insufficient to meaningfully increase buprenorphine use through the end of 2023."

Prescribers rose 27% across 2023 (42,158 in December 2022 → 53,635 in December 2023). Patients treated did not move. The filing states this in Part 2 and draws the local lesson ("the barrier wasn't the binding constraint") — and then keeps the thesis that the binding constraint is law.

A-2. Contingency management inside the VA: no safe-harbor wall, its own cap, and 1.2% delivery

Corrected 2026-08-10, after first publication of this log. This section was originally headed "no legal wall at all" and argued that the VA faces no incentive ceiling of any kind, which made its 1.2% delivery rate a natural experiment showing that no cap was ever the binding constraint. That is contradicted by the source's own Discussion, which names an incentive cap as the first barrier: "Primary among these [barriers] are incentive caps, which hold incentive distributions to <$600 per calendar year due to tax reporting requirements." The VA has a cap. It is a different cap — IRS tax-reporting-driven rather than OIG safe-harbor-driven — at roughly the same magnitude as the safe harbor ($623 for 2026). The claim is narrowed below rather than dropped, because the narrowed version is still evidenced and still bears on the filing's recommendation: what suppresses CM is caps on incentive size whatever their authority, and the OIG safe harbor is one of several, so repealing it would not have moved the VA. What no longer follows from this data point is "even where no ceiling exists, uptake stays low." A second correction in the same paragraph: 138,280 is the count of patients diagnosed with StUD, not an "eligible" population — VA CM is site-based, and the source never characterises the denominator as eligibility. "Eligible" was this record's word, not the source's, and is withdrawn.

The filing's model of CM is that federal fraud-and-abuse law is what stands between the evidence and the patient. There is one place in the United States where that wall has never stood: the Veterans Health Administration, which rolled out a national CM programme in 2011 and standardised CM documentation across its EHR in 2018. Coughlin et al., Am J Psychiatry 2025;182(11):1016–1023 (accepted manuscript read this pass, PMC12872285):

"We identified 138,280 patients diagnosed with StUD between July 2018 and December 2020."

"Among the population with StUD, 1,698 patients received CM for StUD between July 2018 to Dec 2020, and 136,582 did not receive any CM care."

1.2% of the diagnosed population. In the largest integrated health system in the country, in a programme that had been running seven years by the start of that window, centrally funded, with HHS recommending CM as first-line treatment, and with no safe-harbor problem to solve. The same paper says why in its own framing: "CM implementation has been slow, especially in comparison to treatments for opioid use disorder."

What this data point does and does not establish. It does not establish that ceilings are irrelevant, because the VA has one: the paper's Discussion names "incentive caps, which hold incentive distributions to <$600 per calendar year due to tax reporting requirements" as primary among the barriers. What it establishes is that the ceiling on CM is not specific to the OIG safe harbor — a second, independent authority (IRS reporting) imposes a near-identical cap on a system with no fraud-and-abuse exposure, and a safe harbor cannot touch it. Against the filing's recommendation that reads as follows: the filing named one cap and proposed lifting it as the "cheapest, fastest first move"; the largest CM programme in the country would be unaffected by that move and is capped at about the same level regardless. The keystone role this section originally played — "remove the law and uptake still stays low" — is withdrawn. Steelman A's mechanism claim now rests on A-1 (buprenorphine: permission without volume), A-5 (mobile methadone units: staffing and community resistance, not statute) and A-4 (the staffed-service cost model, sourced to Freese et al. independently of the VA), with the VA supplying the narrower cap-plurality point above.

A-3. CMS has approved CM as a Medicaid benefit in five states, above the safe-harbor cap — one is confirmed delivering it

Corrected 2026-08-10, after first publication of this log. Two overstatements, both now fixed here and on both public pages. (i) Approved ≠ implemented. This section and the pages said CM "is already a covered Medicaid benefit in five states." Kaufman et al. say only that five are approved, and state explicitly that "California is the only state to have confirmed implementing"; their Table 1 footnote warns the figures are "subject to change prior to implementing." Five approved, one confirmed running. (ii) The unit was invented. The pages said the ceilings run "$596 to $1,092 per person per year." The source's column is "Maximum Incentive Approved For" and is per programme — 24 weeks for California and Washington, 12 weeks for Montana. Only Delaware's figure is stated annually ($599/year, raised to $750/year). "Per year" is withdrawn everywhere it appeared. Neither correction changes the arithmetic that follows — Washington's $1,092 is still 1.8× the 2025 safe-harbor cap — but it makes the comparison a per-programme-to-annual one, which is stated rather than smoothed over.

The filing's Part 2 says CM "has been held below its trial-effective incentive level by a stack of federal fraud-and-abuse rules." Kaufman et al., Subst Abuse Treat Prev Policy 2025;20:47 (open access, PMC12486863) — a state-by-state review verified with the pilot programme managers — records the position as of March 2025:

"Eight states have applied for Sect. 1115 Demonstration Waivers to implement pilot CM programs. Five states have been approved (California, Washington, Montana, Hawaii, and Delaware), two are pending approval (Michigan, and Rhode Island) and one state's CM application was denied (West Virginia)."

"The CM programs covered under Sect. 1115 Demonstration Waivers vary in substance targeted … length of program (12–64 weeks), amount of incentives ($596–1092) …"

CMS-approved maxima, from the paper's Table 1 — per programme, not per year, except Delaware's: California $599 over 24 weeks, Montana $596 over 12 weeks, Washington $1,092 over 24 weeks, Delaware $750 a year. Washington's approved figure is 1.8× the 2025 patient-engagement safe-harbor cap ($605) and 1.75× the 2026 cap ($623) — noting that the safe-harbor cap is an annual figure and Washington's is a 24-week one, so the true annualised gap is wider, not narrower. Delaware's went up mid-demonstration — the paper's own footnote: "Delaware originally applied for a maximum incentive amount of $599 but later with the approval of CMS the incentive amount increased to $750 a year for both groups."

So the executable federal path to CM at above-safe-harbor incentive levels was open, approved five times before this filing was written, and confirmed delivering in one of those five. It is not the OIG safe harbor. It is §1115 demonstration authority, exercised state by state — which is the inverse of the filing's H6.1 finding that "every binding constraint examined is federal."

And it is faster than the route the filing recommends. From the same paper: "The length from application to approval was one year for both California and Washington … California's 1115 waiver was approved on July 1, 2022 and began implementation in March 2023, nine months after original approval." Application to delivered service: about 21 months. The federal safe harbor the filing calls the "cheapest, fastest first move" has, over the same period, moved from a May 2026 target to Long-Term Actions with an NPRM target of July 2027 (RIN 0936-AA13, re-checked on reginfo.gov this pass — unchanged).

A-4. What actually gets delivered is a staffed service, not a permission

California's programme, described by its own implementation team (Freese et al., Prev Med 2023;176:107703):

"The Program uses a CM protocol where participants can receive a maximum of $599 over a six-month period, contingent upon 36 stimulant-negative urine test results. Urine tests are conducted using a set of approved, CLIA-waived, point-of-care urine drug tests (UDTs). … all aspects of incentive accounting and distribution are managed electronically via a custom-developed software system. … A significant innovation of the project is the conceptualization of the CM Coordinator, a designated and highly trained and supervised individual responsible for all aspects of CM operation in a specific site."

Thirty-six point-of-care tests per patient, a dedicated trained coordinator per site, a custom incentive-management platform, a UCLA-run training and fidelity apparatus, and a readiness review before any site may launch. That is a staffed delivery model with a per-site fixed cost. The filing scores this architecture 3 on fiscal cost — "near-zero net new cost" — because it priced the rule change, not the service.

A-5. Two more barrier removals, for calibration — and one that partly worked

• Mobile methadone units. A 2021 DEA rule let OTPs operate mobile medication units — a legal barrier removed. Miller et al., Addict Sci Clin Pract 2026;21:51 (PMC13308191) interviewed staff at four adopting OTPs: "Key barriers included community resistance to MMUs, unclear or inconsistent guidance from the Drug Enforcement Administration, and a variety of operational challenges, such as vehicle maintenance and workforce shortages."
• Naloxone OTC — the honest counter-example. FDA approved OTC naloxone in March 2023, and here the legal change did move access. A three-phase secret-shopper study of 174 North Carolina pharmacies (Johnson et al., J Am Pharm Assoc 2026, PMC13288354) found same-day availability up (OR 2.59, p=0.0005) and adjusted mean out-of-pocket cost down from $98 to $56 across the three phases. It also found "availability for both product types was low" and independent pharmacies persistently worse. Partial success, not none.

Steelman A does not claim law never matters. It claims that where a legal change is the whole intervention, the measured effect has been small; where the legal change came bundled with a product someone was already stocking and selling, it moved. CM is in the first category, and the filing scores it as though it were in the second.

S3 — Steelman B: the masthead names the architecture the scorecard ranks last

The claim. Part 1 says alcohol and tobacco kill more Americans than the entire overdose crisis. The scope statement says so too, in method terms: "Alcohol and tobacco are in scope as the incumbent legal drugs and the standing comparators; excluding them is itself a policy choice the analysis must not silently repeat." The scorecard then measures mortality impact as "overdose deaths specifically," which excludes every death the masthead counts, and A9 — prevention/tobacco-playbook transplant, the only row aimed at those deaths — falls to joint last. The filing made the excluding choice its own protocol forbade, inside a single dimension anchor.

B-1. The body count, now at CDC-primary tier

Phase 1 recorded the masthead claim as UNVERIFIABLE because cdc.gov 403s every automated fetch. It is reachable — through NCBI rather than CDC. Both legs, fetched this pass:

• Alcohol. Esser MB, Sherk A, Liu Y, Naimi TS. "Deaths from Excessive Alcohol Use — United States, 2016–2021." MMWR Morb Mortal Wkly Rep 2024;73(8):154–161 (doi 10.15585/mmwr.mm7308a1), read in full from PMC10907037 (CC0): "Average annual number of deaths from excessive alcohol use increased 29.3%, from 137,927 during 2016–2017 to 178,307 during 2020–2021; age-standardized alcohol-related death rates increased from 38.1 to 47.6 per 100,000 population." CDC's Alcohol-Related Disease Impact application, 58 alcohol-related causes. The filing's figure and its "2020–21 average" characterisation are both exact.
• Tobacco. The Health Consequences of Smoking—50 Years of Progress: A Report of the Surgeon General (2014), ch. 12, read from NCBI Bookshelf NBK294316: "The results indicate that cigarette smoking and exposure to tobacco smoke led to at least 480,000 premature deaths annually in the United States." With a vintage caveat the filing does not carry: that estimate is the average annual figure for 2005–2009, published 2014. It is CDC's standing number and is still cited as current, but it is a twenty-year-old exposure period, and adult smoking prevalence has fallen substantially since. The number is real; "kills 4–7 times as many Americans annually" is asserted on a 2005–2009 base.

Against the filing's own verified overdose peak (111,466, 12 months ending June 2023, CDC VSRR), alcohol alone is 1.60× the worst year of the fentanyl era. The masthead survives on the alcohol leg by itself.

B-2. The instrument has the evidence the record never went to get

The re-score dropped A9's evidence cell from 3 to 2 because "§9's tobacco arm was never executed — the record verifies the body counts, not the policy efficacy." That is an honest statement about the record, not about the world. The efficacy evidence exists:

• Holford TR, Meza R, Warner KE, et al. "Tobacco control and the reduction in smoking-related premature deaths in the United States, 1964–2012." JAMA 2014;311(2):164–171 (PMC4056770): "an estimated 8.0 million (credible range 7.4–8.3 million) fewer premature smoking-related deaths than what would have occurred under the alternatives and thus associated with tobacco control," and "an estimated 157 million years … of life saved." Tobacco control is associated with 2.3 of the 7.8-year gain in male life expectancy at 40 (30%), and 1.6 of 5.4 years for women (29%).
• The alcohol MMWR closes with CDC's own instrument recommendation: "Implementation of evidence-based policies that reduce the availability and accessibility of alcohol and increase its price (e.g., policies that reduce the number and concentration of places selling alcohol and increase alcohol taxes) could reduce excessive alcohol use and alcohol-related deaths."

Eight million deaths averted is the largest measured effect of any instrument anywhere in this filing. It is not in the filing, because no workstream was run.

B-3. What the anchor does

D4 reads: 1 = "No credible mechanism to overdose deaths specifically" · 3 = "Direct, evidenced mechanism to overdose mortality." Under that wording an architecture that averts 8 million tobacco deaths scores 1, and an architecture with a plausible-but-unquantified path to overdose death scores 2. The anchor is not measuring value; it is measuring which body count the filing decided to name. Steelman B does not ask for A9 to win. It asks for the dimension to stop deciding the answer before the evidence is read — which is the same defect the filing's own re-score identified in A9's original D4=3 and fixed in one direction only.

S3 — Steelman C: political invisibility is fragility, not durability

The claim. Part 7's sequencing rests on this sentence: "A policy that's politically invisible — like a federal safe-harbor rule — can bank results before any backlash coalition has time to organize." Invisibility is scored as a durability asset (A3 D6 = 3, "low backlash risk"). But a policy nobody can see also has nobody demanding it. On a crowded regulatory agenda, salience is what buys calendar time; the absence of a constituency is the absence of a forcing function.

The record is the test, and the record has already run it. Between 2022 and 2026 the federal government made four legal changes to drug treatment and drug law. Three were politically legible and contested — the X-waiver repeal (statute, CAA 2023), OTC naloxone (FDA approval, March 2023), the 42 CFR Part 8 methadone reform (89 FR 7528, 2024) — and the fourth was the cannabis Schedule III order (91 FR 22714, April 2026), which drew three D.C. Circuit petitions. All four happened. The one change in the filing's own catalogue that is politically invisible — a dedicated CM safe harbor — is the one that has not happened, and its published target has moved away from the present: a May 2026 NPRM in the Spring 2025 agenda, Long-Term Actions in Fall 2025, and NPRM 07/2027 on the most recent agenda entry retrievable (pubId 202510), re-fetched from reginfo.gov this pass. Meanwhile OIG published its statutorily mandated annual safe-harbor solicitation in 2022, 2023, 2024 and December 2025 — four consecutive invitations — and CM did not enter a proposed rule through any of them.

And the vulnerability is asymmetric. The re-score explicitly parked this: "OIG-reversibility is a different vulnerability than backlash and stays in the basis text." It is a different vulnerability, and it is a worse one for an invisible rule. A safe harbor created by notice-and-comment can be narrowed or withdrawn by notice-and-comment; conduct outside it reverts to case-by-case anti-kickback analysis. A statutory change (the X-waiver repeal) or a scheduling order with an organised industry behind it (cannabis) is defended by the people who fought for it. A safe harbor nobody has heard of is defended by nobody. Low backlash risk and low durability are the same fact viewed twice.

The comparison the filing draws is also mis-specified. Part 7 contrasts CM's invisibility with Oregon's and San Francisco's reversals — but those were ballot and ordinance policies reversed by electorates. The relevant comparison class for a federal safe harbor is other federal safe harbors, and there the record shows a rulemaking that has not reached a proposed rule across four consecutive annual solicitations.

S4 — Adjudication

A — what it won and what it did not

Won. The mechanism claim. "The binding constraint on American drug policy is not evidence — it's law" does not survive contact with the filing's own evidence base plus the VA and §1115 records. Of the three barrier removals with measured follow-through, buprenorphine's produced permission without volume, mobile methadone units produced implementation friction rather than scale, and only naloxone's — the one attached to a product pharmacies already stock and sell — moved access. The treatment the filing built its recommendation around is delivered to 1.2% of the diagnosed population in the one system that never faced the safe harbor — though, per the A-2 correction, that system has an IRS-driven cap of its own, so this leg now shows cap plurality rather than cap irrelevance. The correct statement is narrower and still interesting: law is a binding constraint on some architectures and a non-binding one on others, and the filing's method — score legal executability as a dimension — cannot tell them apart, because it asks what an agency may do and never asks what a clinic can staff.

Won. The sequencing claim. The filing's "cheapest, fastest first move" is a federal safe harbor with a July 2027 NPRM target. The evidenced fastest route to delivered CM is a state §1115 amendment: about a year to approval, nine more months to service, five states approved and California confirmed delivering since March 2023, CMS-approved incentive maxima up to $1,092 per 24-week programme. Part 7 and ws11 Tier 1 name the wrong instrument.

Did not win. Whether CM is worth doing. The steelman was built expecting to find CM's mortality case thin — the filing itself conceded (via ws13) that the trial base is pre-fentanyl and a 2026 meta-analysis of 26 trials found no overall retention effect. It found the opposite. Coughlin et al. (Am J Psychiatry 2025;182(11):1016–1023) is the first real-world mortality evidence for CM and it is favourable: 1,481 VHA patients with StUD who received CM versus 1,481 matched controls, 41% lower one-year all-cause mortality (aHR 0.59, 95% CI 0.36–0.95). That strengthens A3's mortality cell — the one cell where the filing was under-confident. A steelman that only ever finds against the filing is not a steelman; this one moved a cell in the filing's favour, and it is recorded as such.

Did not win. The incentive-magnitude complaint, cleanly. Rash et al., JAMA Psychiatry 2025;82(9):940–945 give the first data-driven benchmark: from protocols with medium-to-large effects, $128/week for voucher protocols and $55/week for prize protocols — $1,536 and $660 respectively over the standard 12 weeks. Against the voucher benchmark the filing's "held below trial-effective level" is right and the gap is large. Against the prize benchmark — which is the design California and Washington actually run — the $623 safe harbor is within 6% of the effective figure, SAMHSA's $750 grant allowance clears it, and Washington's $1,092 exceeds it by 65%. The withdrawn claim was "capped by an anti-kickback rule"; the surviving claim should be that voucher-magnitude CM is not shelterable under current authority, which is narrower than what the filing said and narrower than what this steelman first assumed.

And the A-2 correction moves this leg further in the filing's favour. The VA's own cap — <$600 per calendar year, imposed by tax-reporting rules — sits below the $660 prize benchmark and at 39% of the $1,536 voucher benchmark. So the largest CM programme in the country is capped below the trial-effective dose too, by an authority the filing never named and a safe harbor cannot reach. The filing's instinct that incentive magnitude is constrained by law was better than the specific law it identified.

B — what it won and what it did not

Won. That the anchor decides the outcome. D4's "overdose deaths specifically" wording assigns A9 a 1 no matter how strong its instrument is, and the filing's own scope statement forbids the exclusion the anchor performs. The re-score noticed this and corrected A9's D4 downward while recording the frame problem as a caveat — an asymmetry that should be stated as such.

Won. That the record never gathered the instrument's evidence. Holford 2014's 8.0 million averted deaths and 157 million life-years is the largest measured policy effect available to this filing and appears nowhere in it, because §9's tobacco arm was never executed and §10 then scored the row on the resulting absence. The filing's own anchored scale rewards research effort here, which is the same defect flagged house-wide in the elder-care re-score.

Did not win. A rank change. Even scored generously (D1 3, D4 3), A9 reaches 16 of 24 — mid-board, behind A5, A10, A11 and A3. And two of A9's low cells are honestly low under any frame: D7 (liberty — the instrument is taxation and restriction, which is coercive by construction) and D8 (speed — tobacco control's mortality dividend is measured over decades, not years). A9's last place is an artefact; a top-three finish is not available to it either.

C — what it won

Unqualified. Every legally visible change in the filing's own catalogue landed within the period studied; the invisible one did not, and receded. The durability score and the sequencing rationale both rest on the reading the record falsifies. The strongest surviving objection to Steelman C is that four observations are a small sample and the CM rule's slippage may reflect HHS agenda capacity rather than salience — true, and it does not rescue "can bank results before backlash organizes," which requires the rule to exist first.

S5 — Scorecard sensitivity

This is a sensitivity analysis, not a re-score. Per M6 and the same reasoning the Phase 1 log applied, cells are not patched outside a structurally blinded pass. The blinded re-score that ws10-findings records as owed is still owed, and nothing below discharges it. What follows is the arithmetic a reader can check, under three explicitly stated readings of the same reconciled matrix (ws10-rescore-log.md).

Reading 1 — as published (v3 reconciled). A3 20 · A5 18 · A10 18 · A11 18/7 (mean 2.571) · A7 17 · A2 16 · A8 16 · A1 15 · A6 14 · A4 13 · A9 13.

Reading 2 — Phase 1's three contradicted cells applied. A3 D2 3→2 (a safe harbor is rulemaking, but the anti-kickback statute survives it and §1320a-7d(a)(1)(B) requires the Secretary to act in consultation with the Attorney General); A3 D8 3→1 (NPRM target 07/2027 is a multi-year lag); A1 D2 1→2 (§823(h) requires a separate registration, not OTP exclusivity). A5 18 · A10 18 · A11 18/7 (mean 2.571) · A3 17 · A7 17 · A2 16 · A8 16 · A1 16 · A6 14 · A4 13 · A9 13. → A3 falls from sole first to joint fourth. Harm-reduction scale-up and the enforcement steelman tie at the top, and the status-quo comparator beats both on per-dimension mean.

Reading 3 — Phase 2's evidence added, A3 scored on the route that actually exists. From Reading 2, plus: A3 D8 1→2 (the §1115 route runs ~21 months application-to-service, an anchor-2 lag); A3 D4 2→3 (Coughlin 2025's real-world mortality association); A3 D3 3→2 (a CM Coordinator per site, 36 point-of-care tests per patient, an incentive platform and a fidelity apparatus is not "near-zero net new cost"); A9 D1 2→3 and D4 1→3 under an all-drug-attributable-mortality reading of D4. A3 18 · A5 18 · A10 18 · A11 18/7 (mean 2.571) · A7 17 · A2 16 · A8 16 · A1 16 · A9 16 · A6 14 · A4 13. → A3 returns to the top — as one of a three-way tie, still behind the do-nothing comparator on per-dimension mean, and only when the recommendation is restated as state Medicaid coverage rather than a federal safe harbor. A9 leaves last place.

What moves and what does not. A3's first place exists in exactly one of three defensible readings — the uncorrected one. Under both readings that incorporate work done since publication, "one clear first move" is not what this board says. What survives all three readings is the comparator: A11, status quo drift, is at or above every architecture on per-dimension mean in every reading. The filing reports this in a parenthesis. It is the scorecard's most stable result.

A reproducibility finding, recorded rather than adjudicated

The whitepaper says the architectures were "ranked under four objective weightings" and the sources page says CM "ranks first or tied-first under 3 of 4 weightings." The weights are not in the committed record. ws10-findings.md names the four scenarios; neither it, nor ws10-rescore-log.md, nor ws11-findings.md gives a weight vector, and no file contains four ranked lists. Only the equal-weight sums are reproducible, which is why every reading above is computed at equal weights. Deviation #13 already discloses that the weights were "set by the synthesizing session's judgment, not pre-registered"; what it does not say is that they were never written down, so the "3 of 4" claim cannot be checked by a reader or reproduced by a re-scorer. Phase 1 marked these claims CONFIRMED against ws10 — correctly, as descriptions of what the record says. This pass records the separate fact that the record does not contain enough to verify them. Verdict: UNREPRODUCIBLE, owed to the same blinded re-score, which should publish its weight vectors.

What the steelman failed to overturn

Recorded per S4's first clause — a conclusion that has beaten a real challenge should say what it beat.

1. CM's clinical case got stronger, not weaker. The steelman went looking for the trial base's weakness and found the first real-world mortality evidence, in the filing's favour.
2. Part 4's enforcement finding held. No evidence surfaced that supply-side enforcement moves price or availability; A10's Σ18 already reflects a genuinely-built steelman, and this pass found no basis to raise or lower it.
3. Part 3's "the decline nobody caused" held, in its corrected multi-cause form. Nothing here attributes the 2023–26 decline to policy.
4. The masthead held, and is now stronger than when it was published — Phase 1 could not reach the CDC figures and marked them unverified; this pass read both at agency-primary tier through NCBI. The alcohol leg alone carries the claim.
5. The two-precedents-one-failure-mode reading of Portugal and Oregon held. Steelman C looked for a durability counter-case and found none; ws12 had already closed Switzerland against the filing.

Sources fetched this pass

All read as full text or publisher-deposited record; none taken from a search summary. Cached with provenance headers under method/sources/:

• cm-delivery-pathways-and-state-medicaid.md — Kaufman 2025 (PMC12486863), Freese 2023 (Prev Med 176:107703), Freese 2024 (JSAT 167:209513), Coughlin 2025 (AJP 182(11), PMC12872285), Rash 2025 (JAMA Psychiatry 82(9)), RIN 0936-AA13 re-check.
• legal-barrier-removal-natural-experiments.md — Chua 2024 (NEJM 390(16), PMC11103581), Johnson 2026 (JAPhA, PMC13288354), Miller 2026 (ASCP 21:51, PMC13308191).
• cdc-alcohol-and-tobacco-mortality-via-ncbi.md — Esser 2024 (MMWR 73(8), PMC10907037), 2014 Surgeon General's Report ch. 12 (NBK294316), Holford 2014 (JAMA 311(2), PMC4056770).

Retrieval note worth carrying to the other filings. Phase 1's largest single gap — 185 UNVERIFIABLE claims, concentrated on cdc.gov's bot block — is partly an artefact of the route, not the source. MMWR is deposited in PMC (often CC0) and Surgeon General reports are on NCBI Bookshelf; eutils.ncbi.nlm.nih.gov and ncbi.nlm.nih.gov/books/ serve both cleanly over curl and are the way around the cdc.gov 403. NIH-funded author manuscripts of paywalled articles (Chua's NEJM letter, Coughlin's AJP paper) are also in PMC. Every filing in this series with CDC-attributed figures should re-try them this way before recording UNVERIFIABLE.

Consequences applied on this branch

Per the corrections protocol, in dependency order — findings file → anchor table → scorecard basis → both public pages.

1. ws04-findings.md — Chua's figures entered at primary tier; the VA 1.2% delivery rate and the five-state §1115 record added; H4.2's implication generalised as the steelman requires.
2. ws06-findings.md — H6.1's "every binding constraint is federal" narrowed: the executable CM path is state §1115 demonstration authority.
3. ws11-findings.md — Tier 1 item 1 re-specified from the federal safe harbor to state Medicaid coverage; the invisibility rationale corrected.
4. ws10-findings.md — the S5 sensitivity readings recorded as owed to the blinded re-score, with the weight-vector reproducibility gap named. No cell re-scored here.
5. research-inquiry.md — anchor 10 (previously blank) filled at primary tier; anchor 9 extended with the state-Medicaid route and the Rash magnitude benchmark.
6. site/drugs/index.html — Part 1, Part 2, Part 6, Part 7 and the honesty box.
7. site/drugs/sources/index.html — §4, §6, §10, §11, receipts and document status.
8. deviations-log.md — entry 23.