Protocol & scope
United States psychoactive drug governance — overdose mortality, addiction and the treatment system, enforcement and criminal justice, and legal-market design. Imports the gubment method (M1–M9) in full: two-source rule, root-tracing for every claim, pre-registered hypotheses, and a deviations log that stays visible rather than getting edited away.
11 workstreams executed as first passes on 2026-08-03: baseline data, the overdose system, treatment economics, enforcement, federalism, market actors, civil society, two root-traced precedents (Portugal, Oregon), an architecture scorecard, and a sequencing plan. A red team pass followed, applying four confidence-calibration amendments before publication.
Anchor table — priors, stated before evidence, verified after
Every anchor below was written down as an unverified guess before research began, so it could be broken. ★ rows were verified in Phase 0; the rest were verified during full execution or remain queued.
| # | Anchor (unverified prior) | Verified value & delta |
|---|---|---|
| 1 ★ | US overdose deaths peaked ~110–112k/yr (2022–23), fell ~25% through 2024–25 | Peak was higher and later (~112–114k, 2023, not 2022–23); cumulative decline is ~37–38%, not ~25%. "Largest decline on record" independently confirmed (Post et al., JAMA Network Open 2025). Corrected 2026-08-10, then corrected again the same day: the June 2025 apparent reversal was a CDC forecasting-model artifact, confirmed at primary tier — Post et al., AJPH 2026;116(5):591–593 (PMC13066679, open access) states "Instead, the anomaly was a model artifact" and that a "second revision released in August 2025" clarified "that the January 2025 'spike' was an artifact." An earlier pass this day withdrew the claim after reading only the title and wrongly assuming the body was paywalled; that withdrawal is reversed. The citation-year fix (AJPH 2026, not 2025) stands. Peak/decline independently re-derived from CDC VSRR provisional counts: 12-month peak 111,466 (June 2023); −37.3% through Dec 2025; −39.4% through Feb 2026. |
| 2 ★ | "Only ~1 in 10 with SUD receive treatment"; perceived need is single-digit % | The "1 in 10" figure is a decade-stale 2015 root (specialty-facility-only, DSM-IV) SAMHSA itself says isn't comparable to current data. 2024 rate is 12.4–19.3% depending on denominator. Perceived need (3.3–6.6%) replicates strongly and is not an artifact of definition. |
| 3 ★ | Federal drug budget ≈$40–45B/yr, ~50/50 supply/demand split; historically ~2:1 supply | Total confirmed ($44.2–44.5B FY23–25); current split corrected to ~55/45 demand-majority. Historical ~65–71% supply-side (1990s–2000s) confirmed exactly. State/local enforcement spending: no unified accounting exists, only a single 1990–91 Census survey, never repeated. |
| 4 | Drug offenses ~13–15% of state prisoners, ~45% of federal; state systems ~7× federal population | Not attempted in Phase 0; verified in full execution — see §5 findings below. |
| 5 ★ | Portugal 2001: deaths and HIV fell post-decrim; attribution to decrim alone is weaker than cited | Confirmed, and worse than the seed anchor implied. HIV decline (~1,287→16 new diagnoses) strongly corroborated. Death data is genuinely incompatible across sources. The dominant US pro-decrim citation (Greenwald/Cato 2009) is confirmed funded advocacy, not scholarship. |
| 6 | Oregon M110: deaths rose post-2021, but synthetic-control studies attribute little to the policy vs. fentanyl timing | Largely confirmed — 2 verified peer-reviewed studies (Joshi 2023, Zoorob 2024) find no detectable effect. Corrected 2026-08-10: previously "3 of 4"; the third could not be located in any index and its reported December-2019 inflection conflicts with Zoorob's own changepoint result (H1 2021). One dissent (Spencer 2023) finds a real effect but attributes it partly to the missing treatment-funding buildout, which didn't disburse until 16 months after decriminalization. |
| 7 ★ | France 1995: office-based buprenorphine cut overdose deaths ~79–80% in four years | Single-root and imprecisely sourced (Auriacombe et al. 2004, underlying dataset unverifiable). Durability contested: French GP buprenorphine prescribing fell 43.8% 2009–2015, and methadone — not buprenorphine — is now France's leading substitution-drug-death cause. |
| 8 | Youth cannabis use did not rise in legalization states | Queued for future §9 execution — not yet verified. |
| 9 | Contingency management capped at ~$75–599/yr by anti-kickback rules | Prior broken on mechanism (corrected 2026-08-10). There is no anti-kickback cap. OIG's own 2020 rule states "there is no OIG-imposed $75 limitation on contingency management program incentives," and the $75 nominal-value guidance runs to the Beneficiary Inducements penalty, not the anti-kickback statute (85 FR 77684, 77791–92). The $75 that bound programs was SAMHSA's own grant condition, lifted to $750 in Jan 2025. The real regulatory ceiling is the patient-engagement safe harbor, 42 CFR 1001.952(hh)(5): $500 base indexed to $623 for 2026, in-kind only, value-based arrangements only. |
| 10 | Alcohol deaths ≈178k/yr, tobacco ≈480k/yr — both exceed peak overdose deaths | Confirmed — see §2 findings below. Alcohol ~178,307/yr (2020–21 avg); tobacco 450–480k/yr. |
| 11 | US methadone dispensable only via ~2,000 OTP clinics; peer countries use pharmacies | Confirmed as to effect; narrowed 2026-08-10. The 2024 reform is verified at primary tier (89 FR 7528) and did leave OTP-only dispensing alone. But §823(h) requires only a separate annual registration on standards the Secretary sets — it never names OTPs or caps their number. The clinic-only structure is 42 CFR Part 8 plus DEA registration standards, not the statute. Asked in 2024 to allow pharmacy dispensing, SAMHSA claimed no statutory bar. |
| 12 | Opioid settlements ~$50–57B/18 years, minority of funds publicly itemized | Confirmed as a genuine moving target ($50–58.6B depending on tracker) — see §7 findings below. Governance is self-policed; only 12 states commit to detailed public reporting. |
| 13 | Retail heroin/cocaine real prices fell for decades while purity rose | Confirmed — see §5 findings below. Real prices fell ~80% since 1981 through the exact decades of maximal enforcement spending. |
Workstream findings
Eleven workstreams, each executed as a first pass with a full source register and two-source-rule discipline. Full writeups (source-by-source, with confidence ratings) are in each workstream's own findings file in the repository.
§2 · BaselineAlcohol and tobacco dwarf the overdose crisis; treatment spending is a genuine unresolved dispute
Verified at primary tier 2026-08-10: alcohol 178,307/yr (2020–21 average, CDC ARDI across 58 causes — Esser et al., MMWR 2024;73(8):154–161) and tobacco "at least 480,000"/yr (2014 Surgeon General's Report, ch. 12), against a verified overdose peak of 111,466 — so alcohol alone is 1.60×. Phase 1 could not reach either figure; cdc.gov 403s automated fetches, and Phase 2 found the block is on the route, not the source (MMWR is deposited in PubMed Central, Surgeon General's Reports are on NCBI Bookshelf). Caveat now carried: the 480,000 figure is the average for the 2005–2009 exposure period, so "4–7×, annually" rests on a twenty-year-old base. Rural treatment deserts are stark: only 6.9% of rural counties had a Medicare-enrolled opioid treatment program in 2022. Total US SUD treatment spending surfaced a ~3× unreconciled disagreement between two credible source lineages ($13.1B vs. ~$34–42B for comparable years) — reported as an open dispute, not resolved by picking a side.
§3 · The overdose systemPolysubstance deaths are now the largest category; the decline's likely causes are outside any policy lever
Opioid+stimulant polysubstance deaths (43.1%) are the single largest death category. The two studies that actually decompose the 2023–25 decline quantitatively point to supply-side/exposure-side mechanisms (a fentanyl supply shock, a shrinking at-risk population) — not naloxone distribution or treatment expansion, which have real but secondary, unisolated contributions.
§4 · Treatment economicsThe X-waiver natural experiment: legal deregulation alone didn't move patient access
2023's buprenorphine prescribing-waiver repeal produced real prescriber growth but flat patient-level access, confirmed by two independent studies — provider willingness/payment binds, not law. Corrected 2026-08-10: what SAMHSA raised 10× in January 2025 was its own grant condition, not a federal anti-kickback cap — no such cap exists (85 FR 77791–92). It was not a rulemaking, and it binds only SAMHSA grantees. Steelmanned 2026-08-10: CMS has approved contingency management as a Medicaid benefit under §1115 in five states (California, Washington, Montana, Hawaii, Delaware) at maxima of $596–$1,092 per programme — 24 weeks in CA/WA, 12 in MT; only Delaware's $750 is an annual figure — with Washington's 1.8× the safe-harbor cap. Only California is confirmed to have implemented, since March 2023. And the VA delivered CM to 1.2% of the patients it diagnosed with stimulant use disorder (1,698 of 138,280, July 2018–December 2020) seven years into a national programme running since 2011. Narrowed on re-check: the VA faces no safe-harbor problem but does face its own incentive cap — Coughlin et al.'s discussion names "incentive caps, which hold incentive distributions to <$600 per calendar year due to tax reporting requirements" as primary among the barriers — so this is evidence that caps on incentive size bind wherever they originate, not that CM stalls where no cap exists. The X-waiver lesson still generalises on delivery: a staffed service is not a permission.
§5 · Enforcement & criminal justice13% of state prisoners are there for drug offenses, and falling; the steelman survives partially
State prisons (7.2× the federal population) hold drug offenders at 13%, down 46% since 2007 — "mass incarceration" is not mainly a drug-war artifact where 89% of prisoners are held. The 40-year price/purity paradox (real prices fell ~80% since 1981 through maximal enforcement spending) is the strongest evidence in the filing against supply-side spending's effectiveness. The enforcement steelman survives partially: focused deterrence has real RCT support for violence, not drug markets; HOPE-style supervision's flagship result failed a 4-site replication.
§6 · Federalism & legal mechanicsDrugs' binding constraints are mostly federal — the cleanest rulemaking-only lever in the filing
Every binding constraint examined (CSA scheduling, the methadone OTP monopoly, the Medicaid IMD exclusion, the anti-kickback ceiling) is federal, not state/local — the inverse of the housing filing's finding. Cannabis rescheduling moved further than expected (a narrow Schedule III order, April 2026, under active D.C. Circuit litigation) but isn't federal legalization. Corrected 2026-08-10: the "payment cap" framing and the "process already underway" claim both fail against primary text. OIG disclaims any CM cap; the November 2024 document cited was OIG's statutorily required annual safe-harbor solicitation (89 FR 93545), which does not mention contingency management; and the actual CM proposal (RIN 0936-AA13) is on the Unified Agenda under Long-Term Actions with a July 2027 target. The cannabis finding, by contrast, was confirmed at primary tier this pass (91 FR 22714; three D.C. Circuit petitions docketed). Narrowed again 2026-08-10 by the steelman: the constraint is federal but the venue is not. CMS has approved contingency-management coverage under §1115 in five states, so "no state or local government can independently fix any of them" fails for the one item this filing recommends — states have been fixing it, one demonstration at a time, since 2022.
§7 · Market actors & the moneyWeak settlement-fund governance; illicit cannabis markets persist at 50–60% even post-legalization
Opioid settlement governance is self-policed — only 12 states commit to detailed public reporting, and $61M+ went to law-enforcement equipment in 2024 alone. California and Washington's legal cannabis markets remain 50–60% illicit, tied explicitly to tax and access costs. Buprenorphine pricing shows a 14–50× disparity between pain and OUD formulations of the identical drug.
§8 · Civil society & coalitionsHarm-reduction vs. recovery-first is a rupture within the anti-punishment coalition, not left vs. right
Fault-line map of harm-reduction orgs, the recovery/abstinence community, law enforcement associations, and prevention groups, with funding sources and a data-vs.-advocacy classification for each. Both Oregon's and San Francisco's 2020-era harm-reduction-forward policies were reversed within 2–4 years by locally organized, donor-backed coalitions — independent of how the underlying mortality data actually resolved.
§9 · PrecedentsPortugal and Oregon, root-traced
Portugal's HIV win is real and strongly corroborated; its death data is genuinely incompatible across sources, and the most-cited US pro-decrim source is confirmed advocacy. Oregon is a rarer case — a genuine multi-study peer-reviewed dispute, not a single-advocacy-root problem — where 2 verified studies find no detectable overdose effect (corrected 2026-08-10 from "3 of 4") and the one dissent locates part of its estimated harm in the same missing-treatment-funding pattern Portugal shows.
§10 · Architecture scorecardContingency management is the only architecture that wins under almost every objective
Eleven candidate architectures scored on 8 anchored dimensions, ranked under 4 objective weightings (mortality-first, order-first, liberty-first, equal). Contingency-management legalization ranks first or tied-first under 3 of 4 weightings — the strongest trial evidence, a rulemaking-only legal path, near-zero cost. Corrected 2026-08-10: the "already underway" basis is withdrawn, and the two cells resting on it (legal executability, speed) are owed a re-score — treat the first-place finish as unsettled. Also corrected: "supply-side modernization ranks last under every weighting" was withdrawn by the 2026-08-04 re-score and should not have survived here; the row now carries a split verdict and the bottom is the treatment-quality and prevention rows. Decriminalization and the enforcement steelman show the largest rank instability, which is the finding, not a flaw. Sensitivity published 2026-08-10: totted up at equal weights under three readings of its own cells, this board puts CM first only in the uncorrected one (20). Apply Phase 1's contradicted cells and CM falls to fourth (17) behind harm reduction and the enforcement steelman (18 each); add Phase 2's evidence and it returns to a three-way tie at 18, with the prevention row leaving last place at 16. Status-quo drift beats every architecture on per-dimension mean in all three. And the "3 of 4 weightings" claim is unreproducible — no weight vector appears in any committed file.
§11 · SequencingThe cheap fix moves first, precisely because nobody's watching it
Three-tier rollout: rulemaking-executable items (CM safe harbor, IMD waivers, settlement-spending conditions) can move within a year; statute-requiring items (methadone deregulation, cannabis banking) queue longer. Public-order political durability runs on a faster reversal clock (2–4 years) than the legal-mechanics clock — front-loading low-visibility rulemaking items banks results before backlash coalitions organize. Both halves corrected 2026-08-10. Tier 1's first item is re-specified from the federal safe harbor (July 2027 target, nothing pending) to state Medicaid §1115 coverage — the route five states have already been approved for and California has confirmed implementing, ~1 year to approval and 9 months more to service. And the invisibility rationale ran backwards: every politically visible federal change in this filing's own catalogue landed between 2022 and 2026, while the invisible one receded across four consecutive OIG solicitations. Low backlash risk and low durability are the same fact seen twice.
Deviations log
Every departure from the protocol as originally written, logged with its reason and effect on findings — per the method, this stays part of the record, not an appendix to it. 23 entries (corrected 2026-08-10 — this page said 16 and the whitepaper said 14; entry 22 records a Phase 1 correction that was itself wrong and has been reversed, and entry 23 records the Phase 2 steelman); the fuller table lives in the repository.
| # | Deviation | Effect on findings |
|---|---|---|
| 1–2 | Phase 0 anchor verification delegated to parallel subagents; mid-run, three were killed and relaunched on a model-selection instruction | No effect — relaunched briefs were verbatim identical; already-completed work was preserved rather than wastefully re-run |
| 3, 6, 7 | Access limitations: cdc.gov blocked direct fetches (row 1), Portugal's two most load-bearing primary papers returned corrupted PDFs (row 5), France's root paper was paywalled (row 7) | Recorded as residual verification gaps in each anchor row rather than smoothed over; confidence capped at moderate accordingly |
| 4–5 | Could not verify AUD/DUD treatment-receipt split (row 2) or the exact demand-majority crossover fiscal year (row 3) | Recorded as "could not verify," not estimated |
| 8 | Non-starred anchor rows (4, 6, 8–13) not attempted in Phase 0 | Left blank, not guessed; closed during full execution |
| 9 | §5–8 run as a second parallel batch; §10–11 synthesized directly by the primary session rather than delegated | §10/§11 are integration tasks across all prior workstreams — synthesized by one author rather than fanned out, consistent with the childcare/housing pattern |
| 10 | Workforce sub-section (§4) rests on a 2037 counselor-shortfall figure and a burnout rate that couldn't be traced to a primary source and were internally inconsistent | Flagged explicitly as the weakest-evidenced part of §4 |
| 11 | §6 found cannabis rescheduling moved substantially further (a finalized narrow Schedule III order, April 2026) than the protocol's seed language assumed | A genuine, dated finding — the record must date-stamp any claim about cannabis's federal legal status given active litigation |
| 12–13 | Red team pass and the whitepaper build were both run by the same session that produced the underlying research | Same single-analyst limitation as GBMT-1's own red team disclosed; an independent second scorer remains a standing requirement before the scorecard is presented as fully settled |
Red team
Five attacks on the §10/§11 synthesis, each answered and, where it landed, absorbed into the record rather than defended against.
Attack 1 — the scorecard's weights were set by the same session that scored the cells
Cannot be fully rebutted from inside. Anchored scales and cited cells reduce but don't eliminate the risk. An independent second scorer remains a standing requirement before the ranking is final.
Attack 2 — contingency management's top ranking may be a dimension-choice artifact, not a discovery
Partially lands — and the 2026-08-10 steelman found the specific cell. The dimension set structurally favors narrow, fast, cheap fixes. Restated in the whitepaper as "the clearest first move," not "the recommendation." The steelman went further and named the mechanism: the mortality dimension is anchored to "overdose deaths specifically," which scores the alcohol-and-tobacco architecture at the floor regardless of its evidence — inside a filing whose masthead is that alcohol and tobacco kill more, and whose scope statement says excluding them "is itself a policy choice the analysis must not silently repeat." The re-score noticed the frame problem and corrected that row downward only. Rescored generously it reaches mid-board, not the top — last place was the artefact.
Attack 3 — if deaths are already falling on their own, does the whole scorecard exercise matter?
Partially lands. Mortality-focused architectures must be read against the status-quo counterfactual; architectures targeting criminal-justice footprint or market design don't depend on which decline theory is correct.
Attack 4 — excluding HOPE from the enforcement steelman may itself be a tilt against enforcement
Does not land. The exclusion is evidence-based: the four-site replication RCT found no advantage over ordinary probation, directly contradicting the single-site result HOPE's reputation rests on.
Attack 5 — Portugal and Oregon share one failure mode; is the "fund treatment fast" lesson unfalsifiable without a positive case?
Lands, and the gap has since closed against us. Switzerland was root-traced on 2026-08-06. It is not a decriminalization case: the Swiss model is a four-pillar framework that retains enforcement, and its exceptional intervention is supervised heroin-assisted treatment for a treatment-refractory population of ~1,700 (about 8% of Swiss heroin users). No spending-effect estimate exists. The trace refutes Switzerland as a positive counterfactual, so the "fund treatment fast" lesson still rests on two cases sharing one failure mode.
Phase 1 verification — every claim, checked against primary law
On 2026-08-10 this filing was re-examined under the series' Verification Protocol by a session that did not author it, did not run its red team, and did not run its blind re-score. Its own citations were treated as claims to be re-derived, not as verification. Coverage is proven by construction: 393 claims extracted, 393 verdicts recorded, zero unaddressed.
| Verdict | Count | What it means |
|---|---|---|
| CONFIRMED | 140 | Re-derived from a primary source fetched in this pass |
| CORRECTED | 23 | Wrong; the correction and its source are recorded |
| OVERSTATED | 34 | Defensible but stated more strongly than the source supports |
| STALE | 11 | Right on the filing date, superseded since |
| UNVERIFIABLE | 185 | No deeper source tier reachable — named, not papered over |
Corrected 2026-08-10: this table previously showed the pass's pre-audit counts (136 / 26 / 186). An adversarial audit the same day reversed one of the pass's own corrections — see the June 2025 item below — moving four claims and giving the post-audit figures above.
The four that matter most. One: the filing's own lead recommendation rested on a legal mechanism that does not exist as described — there is no federal anti-kickback cap on contingency management, OIG says so in its own rule, the $75 figure was SAMHSA's grant condition, and the safe-harbor rulemaking targets July 2027 rather than being "already underway." Two: methadone's clinic-only dispensing is regulatory on top of a narrow registration statute, not "written directly into statute." Three: the Oregon null literature is two verified studies, not three. Four — and this one is a finding against the check, not the filing: the pass withdrew the whitepaper's June 2025 CDC-modelling-artifact claim after reading only the cited article's title and assuming its body was paywalled. The body is open access and says the opposite of what the withdrawal assumed. The claim is restored and confirmed at primary tier; withdrawing a true claim is as much an error as asserting a false one.
Two of these were already sitting in this filing's own research record, found on 2026-08-06 and never carried to the public pages. That gap is the most useful thing this pass found, and it is logged as a deviation rather than quietly closed.
The pass also confirmed the record's single most-flagged unverified claim: the April 2026 cannabis Schedule III order, which the research file had rested on law-firm alerts, is confirmed at primary tier against the Federal Register — scope, dates, the recreational carve-out, the §280E discussion and the June 29–July 15, 2026 hearing window — with three petitions for review docketed in the D.C. Circuit. Two elements were not independently re-verified and should not be read as confirmed: that the three petitions were formally consolidated, and the August 17, 2026 post-hearing-brief deadline, which appears in neither Federal Register notice. Both require authenticated docket access.
verification-log.md — the full ledger, every numbered claim, every verdict, and what could not be reached.
Phase 2 steelman — the strongest case against this filing, built with equal effort
A fact-check asks whether each claim is true. A steelman asks whether the conclusion is right. On 2026-08-10, a third session — independent of the filing, its red team, its re-score and its fact-check — derived this filing's tilt from its own text, then built the best available case against it using the same primary-source standard. The tilt was not hard to find: four of the protocol's own seed hypotheses point the same way, and one of them names the conclusion as "the filing's likely headline shape" before any fieldwork was done.
| Steelman | Verdict | What changed |
|---|---|---|
| DELIVERY BINDS, NOT LAW | Partially survives — and wins on mechanism | The thesis sentence and the Part 7 sequencing are wrong about which lever. The architecture ranking survives, restated |
| THE MASTHEAD ARCHITECTURE IS BURIED | Partially survives | Prevention's last place is an anchor artefact; a published sensitivity, not a rank change |
| INVISIBILITY IS FRAGILITY | Survives | Part 7's rationale and the durability cell behind it both rest on a reading the record falsifies |
The evidence that did the work. CMS has approved contingency management as a Medicaid benefit under §1115 in five states — California, Washington, Montana, Hawaii, Delaware — at incentive maxima of $596–$1,092 per programme (24 weeks in CA/WA, 12 in MT; Delaware's $750 is the one annual figure), Washington's being 1.8× the federal safe-harbor cap this filing calls the binding ceiling. California is the only one confirmed implementing, and has delivered since March 2023. Meanwhile the VA, which faces no safe-harbor exposure and has run a national CM programme since 2011, delivered it to 1,698 of 138,280 patients diagnosed with stimulant use disorder — 1.2%. Narrowed on re-check: the VA is not a no-ceiling case. Coughlin et al.'s own discussion names "incentive caps, which hold incentive distributions to <$600 per calendar year due to tax reporting requirements" as primary among the barriers — a different cap from the OIG safe harbor, at about the same level. The point that survives is that caps on incentive size bind wherever they come from, and repealing the OIG one would not have moved the VA. The delivery cost stands on its own evidence: a trained coordinator per site, 36 point-of-care tests per patient, an incentive platform, a readiness review.
What it could not take down, and what it strengthened. Part 4's enforcement finding and Part 3's corrected "decline nobody caused" both held. And the case for contingency management as a treatment came out stronger: the first real-world mortality evidence, published in 2025, found CM recipients 41% less likely to die within a year (adjusted hazard ratio 0.59) in a matched VA cohort — a cell where this filing had been under-confident. A steelman that only ever finds against the filing isn't a steelman; this one moved a cell in the filing's favour and says so.
steelman-log.md — the tilt derivation, all three steelmen with their sources, the S4 adjudication, and the scorecard sensitivity arithmetic in full.