MED01-P1 — Critique & Verification (v0.1)

Verifier: Claude · Date: 2026-09-23 · Inputs: raw-inputs/chatgpt/MED01-P1-chatgpt-pass-raw.md (24 CAND, 39 refs) and raw-inputs/gemini/MED01-P1-gemini-pass-raw.md (25 CAND, 29 refs), both frozen as received. Method: every PMID/DOI resolved live via NCBI E-utilities and Crossref (identity); PubMed abstracts read for every quantitative clinical claim (support); Crossref update-to checked on all shared DOIs (publication integrity); both researchers’ PubMed count queries re-run verbatim; drift checks D1–D14 applied per candidate claim. Firewall lifted on filing of the second pass; this file is the first document to quote both.

1. Independence disclosures (must weigh before convergence)

Both researchers disclosed that prior project context was visible. ChatGPT also reported the attachments were not mounted and that it fetched the frozen brief and protocol from Drive itself. Neither run was a clean session as the kickoff intended. Consequence: convergence on the discovery chain (1963/64/65 Mechoulam group → 1988 Devane → 1990 Matsuda/Herkenham → 1992 anandamide → 1993 CB2 → 1995 2-AG) is not treated as independent confirmation — these are also the landmarks any model reaches. Convergence carries weight only where the two passes reached the same specific datum by different routes, and divergence is treated as the more informative signal in this round. Recommendation for P2 in §7.

2. Reproducible count (brief §3A) — verifier re-run 2026-09-23

Findings: Gemini’s total and every decade figure differ from the verbatim re-run (total −5.4%; 1970s −10.7%; 1960s +21%). PubMed counts drift with indexing, but not by a decade-level 10% in three days, and Gemini gave percentages to one decimal computed on its own figures. Disposition: Gemini’s count is not reproducible as stated; verifier count supersedes (ledgered as verifier-found). ChatGPT’s honesty in reporting no count is the protocol-conformant behavior. The shape — a 1970s surge, a 1980s dip, a 1990s peak exceeding the 1970s — is robust across both strings and is the only count-derived claim that enters the ledger.

3. Check 1 — citation identity

3a. ChatGPT (39 references): 38 exact, 1 wrong-object

• All 34 PMIDs resolve to the stated paper; all DOIs resolve. Volume/page/year exact throughout, including the hard cases (Petro 21:413S; Cronin 21:43S; Hepler 217:1392 no abstract).
• ❌ PMID 35817852 cited for “a later period review indicates dozens of cannabinoids were known by 1980” → resolves to Busquets-García et al., Nat Metab 2022, “Metabolic Messengers: endocannabinoids.” Wrong object; the claim it supported (Q1 minor cannabinoids) was already marked incomplete by ChatGPT, so nothing load-bearing rests on it. Logged.
• ⚠️ Ng & Chang 2022 co-author given as “Chang YK”; Crossref: Nathan Chang. Metadata slip, identity otherwise exact.

3b. Gemini (29 references + in-text citations): 3 identity failures, 2 metadata errors, 3 unresolved

• ❌ Beal 1995 DOI 10.1016/0885-3924(94)00111-7 → 404. Real DOI …00117-4 (ChatGPT’s). Paper is real; wrong identifier.
• ❌ Tashkin 1973 DOI 10.1056/NEJM197308162890701 → resolves to Gordis et al., NEJM 289:331–335 (rheumatic fever care programs). Wrong object. The real Tashkin 1973 paper is NEJM 289:336–341, PMID 4718513 — Gemini’s pages “333–336” are also wrong. Substance (acute bronchodilation, smoked marijuana + oral THC, healthy volunteers and asthmatics) is right.
• ❌ Gill, Paton & Pertwee 1970 “Nature 228(5277):1193–1194” → the indexed paper is Nature 228:134–136, “Preliminary experiments on the chemistry and pharmacology of cannabis” (PMID 5466704). Wrong pages; whether it is the THCV-isolation paper is not confirmed from the record.
• ⚠️ “Mechoulam et al. 1967 JACS 89:4552 — total synthesis” → real paper (PMID 6046550) is “A stereospecific synthesis of (−)-Δ1- and (−)-Δ1(6)-THC.” The total synthesis is Mechoulam & Gaoni 1965, JACS 87:3273 (ChatGPT’s citation). Gemini’s Q1/CAND-01 chronology conflates the two.
• ⚠️ Tashkin 1999 AJRCCM 160(5):1658–1666 → no PubMed record for a Tashkin AJRCCM 1999 paper at that location. Unresolved; treated as unverified.
• ❓ Hall, Room & Bondy 1999 Addiction (the “excised WHO chapter”) → not found in PubMed by author/year/journal. Unresolved. Claussen, von Spulak & Korte 1966 Tetrahedron → not indexed (old chemistry); unresolved, not a failure.
• ✅ Confirmed on Gemini’s side where ChatGPT had nothing: Frytak 1979 (PMID 517882), Tashkin 1987 ARRD 135:209 (PMID 3492159), McGlothlin & West 1968 (PMID 5667203), Stenchever 1974 (PMID 4808863), Nahas 1974 Science (PMID 4271816), De Petrocellis 1998 (PMID 9653194), Fried Ottawa cohort papers (1996–1999).
• Pattern: identical to R01–R07 — Gemini’s substance largely survives; its identifiers and page numbers cannot be transcribed without resolution. Several links were Google-search wrappers.

3c. Publication integrity (§M1, v0.4)

Crossref update-to on all 19 shared DOIs: no retraction, withdrawal, expression of concern, or correction on any. Smith 2015 Cochrane shows new_version (a later edition exists — cite the version used). Date checked: 2026-09-23.

3d. Ng & Chang 2022 (the bibliometric anchor both passes lean on)

Identity exact. Full text was not retrievable through Europe PMC in this pass, so neither researcher’s decade figures from it (ChatGPT: US 33 / 979; Gemini: ~300 / 2,350 / 1,550 / 3,100 and a Scopus TITLE() string) are confirmed from the paper. Until the PDF is read, these are abstract-only-equivalent and capped under §M9. The PubMed re-run independently supports the decade shape.

4. Check 2 — claim support (abstract-level, PubMed)

Not checkable at abstract level (full text or primary document needed): Congressional Record 1969/1970/1974 items (ChatGPT; govinfo PDFs cited with URLs — plausible, to be opened); Federal Register 64 FR 35928 (both; consistent); 54 FR 53767 and the 1988 Young ruling (Gemini gives docket and dates; ChatGPT explicitly did not retrieve the original — Gemini’s specifics are unverified until the FR page is opened); compassionate IND figures (Gemini: 1978 / June 1991 halt / March 1992 closure / 27 approved / 13–15 grandfathered; ChatGPT declined to assert) — these match Danny’s archive claims A18/A20/A33 verbatim, which is a reason for more scrutiny, not less; Gemini’s “>75% of federal research budgets” on harms — no source given; violates §3G’s empirical-basis rule; rejected; Gemini’s “over 90% accuracy” unblinding figure and “24–48 h washouts” — no source; rejected as numbers, kept as qualitative concern.

5. Drift and rule findings

• D5 (route): Gemini §D3/Q10 presents Chang 1979 and Frytak 1979 as smoked-vs-oral comparisons with PK explanation; Chang’s abstract does not support an isolated route contrast, and Frytak compared oral THC with prochlorperazine and placebo (title). Reword to what the studies did.
• D8/D11 (surrogate/overstatement): Gemini Tier 1 on dronabinol and weight; Gemini Tier 1 “efficacy comparable to codeine 60–120 mg.”
• D2 (causal): Gemini G3 wording is careful; Andréasson abstract itself is the causal-language source — ledger will attribute.
• Rule 6a (hindsight): Gemini’s Tier 3 items “disproven by longitudinal testing” / “disproven by biopsies” cite no period source and read as present-day verdicts; ChatGPT’s Tier 3 is correctly phrased as “not identified in this search.” Gemini’s Tier 1 inclusion of the ECS as “established in humans” is a category slip — receptor/ligand evidence was molecular and animal (both passes agree on the evidence; only Gemini’s tier placement is wrong).
• §3C motive rule: Gemini supplies motivating narratives (“anecdotal self-medication prompted trials”; “no approved appetite stimulants before megestrol”) without contemporaneous citation; ChatGPT stayed within sourced statements.
• §3G empirical-basis rule: ChatGPT declined to quantify the harms/therapeutic balance; Gemini asserted “>75%” unsourced and “largest single component” unsourced. ChatGPT conforms.
• §4a database honesty: Gemini’s CAND-24 cites “EMBASE scans” not listed in its search log; its search log lists 9 strings for 29 references. ChatGPT’s log is complete.
• Both: neither flagged Herman 1979’s doubled AE rate; both treat IOM 1999 conclusions correctly as IOM’s.
• Positive divergence worth ledgering: Gemini surfaced Frytak 1979, Tashkin 1973/1987, the Shafer and Le Dain commissions, IOM 1982, the NIH 1997 workshop, the 1988 Young ruling / 1989 Lawn denial, the Randall case, Stenchever/Nahas refutations, and the WHO 1994–97 excised-chapter episode — none in ChatGPT’s pass. ChatGPT surfaced Williams 1976, Petro 1981, Benowitz 1980, Lemberger 1975, Cronin 1981, DMHP 1974, Brewster 1997, the 1969/1970/1974 Congressional Record supply documents, the UK Government’s Dec-1998 response, and NCT00000737 — none in Gemini’s. The union is a materially better map than either pass.

6. Convergence map (for the ledger)

7. Open items before ledger v0.1

1. Open the govinfo PDFs (CR 1969/1970/1974; 64 FR 35928; 54 FR 53767; 57 FR 10499) and the IOM 1999 chapter text — page-level support for the regulatory claims.
2. Retrieve Ng & Chang 2022 PDF for the decade figures and the Scopus string.
3. Primary source for compassionate-IND dates/counts (FDA/PHS record or IOM 1999 pp. 43–44) — until then, Q12 is Inconclusive.
4. Full text for Cunha 1980 (per-patient breakdown) and Hepler & Frank 1971 (any effect size).
5. Correction round to Gemini: Beal DOI; Tashkin 1973 DOI/pages; Gill 1970 pages; Tashkin 1999 location; Hall/Room/Bondy 1999; “1967 total synthesis”; sources for 25–30% IOP, codeine equivalence, >75%, >90%, 24–48 h. Correction round to ChatGPT: PMID 35817852; Chang co-author. Corrections inherit full verification (Rule 9).
6. Completeness probe (§M12b): Cochrane/PubMed for systematic reviews and pivotal trials neither cited — run before ledger.
7. P2 kickoff: deliver brief and protocol as local attachments in a genuinely fresh account/session, never via Drive; require the isolation disclosure to name what was visible. Consider having Danny run the pass from a browser profile with no project memory.