MED01 — Verification Log

Every citation offered by either researcher gets a row, including the ones that fail. Failures stay here permanently; they are part of the audit trail and feed the SOURCE_LIBRARY claim-support cautions table.

Check 1 — Citation identity: the source exists, and authors / title / venue / year / identifier match. Includes publication-integrity status (retracted / withdrawn / expression of concern / corrected — §M1), with the date checked, recorded in the identity column.
Check 2 — Claim support: the source actually supports the statement attributed to it.
Drift: which PROTOCOL-MEDICAL §M7 checks (D1–D14) the attributed claim failed, if any.

Result codes: ✅ pass · ⚠️ pass with correction (state it) · ❌ fail · ⏳ pending

#	Cited as (verbatim from researcher)	Researcher	Check 1 identity	Resolved identity (S-ID, DOI/PMID)	Claim attributed	Check 2 support	Drift flags	Access level	Outcome → claim ID / failure mode	Date
1

Citation failures (summary)

Cited as	Researcher	Failure mode	Notes
		invented paper · wrong DOI · wrong authors · wrong year · citation conflation · review cited as trial · real paper, wrong claim

Verifier self-check (base PROTOCOL v1.4)

Any claim where Claude overturned a researcher is listed here with the evidence for the overturn, so the secondary audit can test the verifier too.

Claim	Researcher said	Claude ruled	Basis

P1 verification run 1 — 2026-09-23 (Claude; NCBI E-utilities + Crossref; abstracts read)

Full narrative in critique.md. Summary rows (identity = Check 1 incl. publication-integrity; support = Check 2 at abstract level unless noted):

#	Cited as	Researcher	Check 1	Resolved	Check 2	Drift	Outcome
1	Ng & Chang 2022 “Chang YK”	ChatGPT	⚠️ co-author = Nathan Chang	DOI 10.1186/s42238-022-00133-0	⏳ full text not retrieved; decade figures unconfirmed	—	metadata slip; §M9 cap until PDF read
2	PMID 35817852 “period review, dozens of cannabinoids by 1980”	ChatGPT	❌ wrong object — Busquets-García, Nat Metab 2022	—	n/a	—	citation failure (non-load-bearing)
3	34 other PMIDs / DOIs	ChatGPT	✅ all exact	see critique §3a	✅ where abstract available	—	pass
4	Beal 1995 DOI …00111-7	Gemini	❌ 404; real …00117-4	PMID 7730690	✅	D8/D11 on weight wording	wrong identifier
5	Tashkin 1973 NEJM 289:333–336, DOI …2890701	Gemini	❌ DOI → Gordis 1973; pages wrong	PMID 4718513, 289:336–341	✅ substance	—	wrong object DOI + wrong pages
6	Gill, Paton, Pertwee 1970 Nature 228:1193	Gemini	❌ pages; indexed 228:134–136	PMID 5466704	❓ THCV claim unconfirmed	—	wrong pages
7	Mechoulam 1967 JACS “total synthesis”	Gemini	⚠️ real paper = stereospecific synthesis; total synthesis = 1965	PMID 6046550 / 14324315	—	—	chronology conflation
8	Tashkin 1999 AJRCCM 160:1658	Gemini	❓ no record found	—	—	—	unresolved
9	Hall, Room, Bondy 1999 Addiction	Gemini	❓ not found in PubMed	—	—	—	unresolved
10	Hepler & Frank 1971 “25–30%”, “11 volunteers”	Gemini	✅ identity	PMID 5109652 (no abstract)	❌ figures not in cited source	6a	unsupported from source
11	“>75% of federal budgets” harms; “>90%” unblinding; “24–48 h washout”; codeine 60–120 mg equivalence	Gemini	— no source	—	❌	§3G, D11	rejected as numbers
12	Chang 1979 as smoked-vs-oral comparison	Gemini	✅ identity	PMID 293141	❌ abstract gives no route contrast	D5	reword
13	Frytak 1979, Tashkin 1987, McGlothlin 1968, Stenchever 1974, Nahas 1974, De Petrocellis 1998, Fried 1996–99	Gemini	✅	PMIDs 517882, 3492159, 5667203, 4808863, 4271816, 9653194, 8809875/9638687/10492386	⏳	—	pass identity
14	Shared core (Sallan, Chang, Cunha, Beal, Andréasson, Munson, Hampson, Devane ×2, Matsuda, Herkenham, Munro, Mechoulam 1995, Sugiura, Herman, Gaoni 1964, Mechoulam-Shvo 1963)	Both	✅ all exact; no retraction/EoC/correction (Crossref, 2026-09-23)	—	✅ at abstract level	Herman AE rate under-reported by both	pass
15	PubMed count query	Gemini	—	verifier re-run 10,630 (vs 11,241); decades 414/3,679/2,516/4,021	❌ not reproducible as stated	—	superseded by verifier-found count
16	PubMed count query	ChatGPT	—	verifier re-run 8,663; 252/2,492/2,153/3,766	n/a (none claimed)	—	verifier-found

Citation failures (summary) — P1 run 1

Cited as	Researcher	Failure mode	Notes
PMID 35817852	ChatGPT	wrong object	non-load-bearing
Beal 1995 DOI	Gemini	wrong DOI	paper real
Tashkin 1973 DOI + pages	Gemini	wrong-object DOI; wrong pages	paper real
Gill 1970 pages	Gemini	wrong pages
“1967 total synthesis”	Gemini	conflation
Tashkin 1999; Hall 1999	Gemini	unresolved	correction round
Hepler 1971 magnitudes; >75%; >90%; 24–48 h; codeine equivalence	Gemini	attributed numbers not in any cited source	correction round

P1 correction round 1 — ChatGPT (2026-09-24)

#	Item	Returned	Check 1	Check 2	Outcome
C1	replacement for PMID 35817852	Turner, ElSohly & Boeren 1980, J Nat Prod 43:169–234 (PMID 6991645, DOI 10.1021/np50008a001) + NIDA Res Monogr 31 (PMID 6775230)	✅ both exact	⏳ “61 cannabinoids” unconfirmed (no abstract)	accepted; capped until full text
C2	Ng & Chang co-author	Nathan Chang; PMID 35610633	✅	full-text read asserted for 33/979 — verifier to confirm	accepted
C3	count	acknowledges 8,663	—	—	noted
Round: 3/3 resolved, 0 new failures. Gemini correction round pending.

P1 correction round 1 — Gemini (2026-09-24) — 18 items; full table in raw-inputs/gemini/MED01-P1-gemini-correction-1.md

#	Item	Outcome
G1	Beal DOI	✅ resolved
G2	Tashkin 1973	DOI ✅; PMID 4719469 wrong-object (DiBona) — real 4718513
G3	Gill 1970	DOI ✅; PMID 5460037 wrong-object (Winans) — real 5466704; THCV priority unconfirmed
G4	synthesis chronology	3 DOIs ✅; ruling: 1965 Mechoulam-Gaoni = first total synthesis; Fahrenholtz 1967 + Petrzilka 1969 added as verifier-confirmed
G5	Tashkin 1999	withdrawn → PMID 9001303 ✅
G6	Hall/Room/Bondy	withdrawn as article; book chapter unverified
G7	Claussen 1966	DOI ✅
G8	Hepler “25–30%”	AJO 1972 DOI and PMID 4647805 both wrong-object (Pfaffenbach; Freilich); magnitude REJECTED — no resolvable source
G9	“>75%”	withdrawn ✅; replacement claim about Ng & Chang not in the text — rejected
G10–11	“>90%”, “24–48 h”	withdrawn ✅
G12	Noyes codeine	DOI ✅; PMID 1097148 wrong-object (Hollister) — real 50159; “20 mg ≈ 60 mg codeine” abstract-only, capped
G13	compassionate IND	HOLD — memoir + IOM page cite; Q12 Inconclusive
G14	54 FR 53767 quote	HOLD — FR page not yet retrieved by verifier
G15	nabilone Canada/UK	HOLD — year now 1981 vs pass 1982; unverifiable numbers
G16	EMBASE	admitted not searched; CAND-24 narrowed
G17	count	accepts 10,630
G18	Chang 1979	reword accepted; PMID 389332 wrong-object (Weibel) — real 293141
Round: 18/18 answered · 5 new wrong-object PMIDs + 1 wrong-object DOI introduced · 8 withdrawals accepted · 4 held.

Verifier self-check additions (P1)

Claim	Researcher said	Claude ruled	Basis
Hepler & Frank 1971 IOP magnitude	Gemini: 25–30%, 11 subjects	not ledgerable	cited letter has no abstract; both replacement identifiers resolve to other papers; verifier-found Cooler 1977 / Green 1998 to be read first
PubMed count	Gemini: 11,241	10,630	verbatim ESearch re-run 2026-09-23; Gemini accepted
Harms/therapeutic balance	Gemini: >75%, then “majority per Ng & Chang”	cannot be quantified	Ng & Chang full text codes funders/countries, not harm vs therapeutic
First total synthesis of Δ9-THC	Gemini: 1967	1965	Crossref/PubMed titles: 1965 “A Total Synthesis…”, 1967 “Stereospecific synthesis…”

Researcher-performance record — P1 (moved from source-library-additions per ChatGPT v0.1 review)

Researcher	Observation	Consequence
Gemini correction round, MED01-P1	5 new wrong-object PMIDs (4719469, 5460037, 4647805, 1097148, 389332) + 1 wrong-object DOI, each an adjacent record to the real paper — caught: Rule 9 re-resolution	Fifth consecutive Gemini correction round to add identity failures (R01, R02, R04, R07, MED01-P1). Never transcribe a Gemini identifier without resolution.
ChatGPT, MED01-P1	PMID 35817852 wrong-object (Nat Metab 2022) for a 1980-era cannabinoid count — caught: run 1	Replaced by Turner 1980 (S-251); ChatGPT’s only P1 identity failure; replacement exact.

Standing rule (base Rule 9, PROTOCOL-MEDICAL §M1): never transcribe a correction identifier without live re-resolution.

§M10a cross-audit — auditors’ own claims checked (2026-09-26)

Auditor	Item	Check	Outcome
ChatGPT	MeSH “Cannabinoids” year introduced 1977	NCBI MeSH D002186 page: “Year introduced: 1977”	✅ confirmed; ledger C001 corrected (was 1971)
ChatGPT	C019 breakdown is in the abstract	efetch PMID 7413719 full abstract	✅ confirmed; verifier’s 2026-09-23 read was truncated — correction #2
ChatGPT	IOM 1999 / CR 1970,1974 / NARA / Canada.ca / NIH 1997 / FR 1992 passages	not reproduced by verifier (Bookshelf blocked; PDFs not opened)	logged as cross-auditor convergence; HOLDs stay
Gemini	C001 counts reproduced live via NCBI (10,630 / 8,663)	matches verifier’s two runs	✅ both-model requirement met on Gemini’s side
Gemini	six “verbatim” synthesis quotations (rows 15, 17, 18, 19, 28, 31; §2.3 commissions)	compared against synthesis v1.0 text	❌ not in the document — auditor quoted its own reconstruction; major finding on the auditor
Gemini	Noyes 1975b as PMID 1097148 (row 14)	PMID 1097148 = Hollister 1975	❌ repeat of a wrong-object PMID already caught 2026-09-24

Verifier self-check additions (cross-audit)

Claim	Verifier recorded	Correct state	Basis
C019 per-patient breakdown	“not in abstract”	in abstract (4/8, 3/8, 1; 7/1)	truncated read; caught by ChatGPT audit — conservative error, disposition unchanged
C001 MeSH year	1971	1977	NLM record; caught by ChatGPT audit

P2 verification run 1 — 2026-10-06 (Claude; NCBI E-utilities, Crossref, ClinicalTrials.gov v2, Health Canada DPD, openFDA)

Full findings in critique-P2.md. Summary:

Canonical counts (verifier, live): 2000–04 4,335 · 2005–09 7,451 · 2010–14 10,316 · 2015–17 8,599 (total 30,701). ChatGPT: unreproducible, not estimated (compliant). Gemini: 5,618 / 7,542 / 9,874 / 8,365 reported despite stated blocked access — none reproduces (§4a breach).
Identity, ChatGPT: all PMIDs/DOIs resolve to named papers; NCT00046722, NCT00681538, NCT02091375, NCT01040910 resolve; DIN 02266121 confirmed. ISRCTN68314063 pending verifier open.
Identity, Gemini (rerun S1 + S2): 27 wrong-object/nonexistent PMIDs (table in critique §3b, with correct PMIDs located by the verifier); 7 nonexistent + 3 wrong-object DOIs; DINs 02269926 / 02268571 nonexistent; two bibliometric papers attributed to wrong authors; five “resolved title copied” titles not the resolved titles.
Integrity: no retraction / expression-of-concern / Crossref updated-by on any resolved identifier (2026-10-06). Koppel 2014 flagged [RETIRED] guideline status.
Regulatory: Syndros NDA 205525 ORIG1 2016-07-01 confirmed; Cesamet NDA 018677 ORIG1 1985-12-26, 2006-05-15 = labeling supplement (not an approval).
Claim support (abstract): ChatGPT figures exact on CAMS, Ware, Abrams, Kerbrat, Naftali, Geffrey, Tramèr. Gemini not supported/contradicted on CAMS THC CI and funder, Ware CI and dosing days, Abrams design and p, Portenoy primary (D11), Whiting spasticity estimate, Koppel scope/wording, Kerbrat counts, Naftali endoscopy (2021 study, hindsight).
Verifier-found: NCT01040910 current primary outcome (CDAI −70 response) differs from the published primary (remission) — registry-history check pending.

Researcher-performance record — P2 run 1

Researcher	Unique PMIDs cited (passes of record)	Wrong-object / nonexistent	Count compliance
ChatGPT	54	0	compliant (not estimated)
Gemini	62	27 (44%)	breach (unreproducible figures reported)

P2 correction round 1 — ChatGPT (2026-10-06)

3/3 resolved, 0 new failures (raw + re-resolution: raw-inputs/chatgpt/MED01-P2-chatgpt-correction-1.md). Verifier-found: ISRCTN68314063 (Ware 2010) was assigned 2006-06-21, after recruitment 2003-08-01 → 2006-01-31 — retrospective registration; registry names both 100-mm VAS and 11-point NRS. Bears on RoB 2 Domain 5 for Ware 2010.

P2 correction round 1 — Gemini (2026-10-06)

55 items; full table in raw-inputs/gemini/MED01-P2-gemini-correction-1.md. Confirmed 38 · withdrawn 2 (incl. item 42: the four canonical counts were “estimated from background training approximations” — fabricated-data finding, repeats P1) · not resolved 1 (Tzadok; verifier has 26800377) · new failures 4 (Rocha 2014 replaced by an unrelated 2008 CINV meta-analysis; Matielo first name; Naftali 2021 pages; Cesamet 2006 labeling supplement called an efficacy-supplement approval) · unverified full-text assertions 6 (CAMS 78%/capsule, Whiting −0.76 and AE ORs, Kerbrat regulatory cohort, Novotna −0.84, Meier 8 IQ points, Geffrey CYP2C19). Access declarations (item 41) put most Gemini candidates under the §M9 abstract-only cap and contradict its own full-text page citations. Pletcher analytic N 5,016 confirmed in PMC full text.

Researcher-performance record — P2 (after correction round 1)

Researcher	Run-1 identifier failures	Fixed in round 1	New failures in round 1	Count compliance
ChatGPT	0	—	0	compliant
Gemini	27 PMIDs + 10 DOIs + 2 DINs	25 PMIDs, all 10 DOIs, DIN	4 (1 substitution, 3 metadata/characterisation)	fabricated (admitted)

P2 record checks — 2026-10-09 (Claude; ClinicalTrials.gov v2, PubMed, NAP reader, Health Canada NOC/DPD, GOV.UK, EMA, openFDA, Google Patents, CourtListener, FederalRegister.gov)

Full table in critique-P2.md §6b–§6c.

Registry matches: 9 of 10 candidate trial→paper matches confirmed by PI/site/product/N; NCT00309413 ↔ Leweke 2012 rejected (paper lists NCT00439634). Schimrigk 2017 (NCT00959218): primary not met (p=0.676), abstract conclusion emphasises safety.
NASEM 2017: Summary conclusions opened via the NAP reader (NCBI Bookshelf still blocked by reCAPTCHA). Both passes’ NASEM statements consistent with the conclusion numbers.
Health Canada: Sativex NOC/c 2005-04-15 (091289) and 2007-08-01 (109461) confirmed; both indications removed 2019-12-12 (retrospective).
MHRA: June 2010 confirmed; day not verifiable. EMA Acomplia: MA 2006-06-19, withdrawal 2009-01-16 confirmed. Epidiolex: NDA 210365 approved 2018-06-25 (P3). FDA orphan-designation dates not verified (OOPD returned nothing usable).
Patents: US 6,630,507 — ChatGPT exact. New finding against ChatGPT: US 9,474,726 claim 1 is CBD for FIRES, not Dravet. Gemini: US 9,066,920 (partial seizures, not Dravet/LGS) and US 8,071,641 (Hadasit/Yissum, diabetes; not GW) mischaracterised.
Legal/regulatory: Craker v. DEA 714 F.3d 17 (1st Cir. 2013-04-15) · 81 FR 53846 (2016-08-12) · 82 FR 14815 (2017-03-23, Syndros Schedule II) confirmed.

Researcher-performance addendum — record checks

Researcher	Official-record items checked	Correct	Mischaracterised
ChatGPT	9 (HC ×2, MHRA month, EMA, US 6,630,507, US 9,474,726, 81 FR, 82 FR, Syndros)	8	1 (US 9,474,726 scope)
Gemini	8 (HC ×2, MHRA month, EMA, US 6,630,507, US 9,066,920, US 8,071,641, Craker)	6	2 (both patent scopes/assignees)

P2 §M10a cross-audit — 2026-10-13

Full record: secondary-audit-P2-verification.md and dispositions.

ChatGPT audit accepted. Confirmed at source by the verifier:
- Wade 2004 primary outcome;
- Pletcher FEV1 and FVC (not the ratio);
- US 8,071,641 = Type II diabetes;
- Wang 96.6% = events within 23 RCTs;
- NCT01898520 published in 2020;
- Devinsky 2016 ages 1–30;
- Meiri response rates;
- Tramèr full text “30 randomised controlled trials”;
- Meier 8 IQ points (n = 23 subgroup);
- UK 16 June 2010 and Spain 20 July 2010.
Convergence, verifier read pending: Whiting corrected spasticity values and adverse-event ORs (JAMA bot-blocked); IGAS cohort; TGA; FDA orphan designation; Novotna −0.84; NCT01040910 protocol; Dutch 2003; ACP/AMA text.
Gemini audit rejected: 22 of 26 sampled source quotations not in the abstracts; 11 of 11 quoted MED01 passages do not exist; 4 wrong PMIDs.
Verifier failures recorded:
- the integrity check omitted ErratumIn (7 sources flagged on rescan);
- the scheduled registry title search was never run (C067 re-screened: 28 therapeutic → 17 matched, 11 unresolved);
- the “comparisons ≠ trials” remark against Gemini was wrong (withdrawn);
- the Kerbrat ruling was overbroad.
Corrections #5–#13 applied to the ledger, synthesis and plain-English pair (corrections.md).

Researcher-performance addendum — cross-audit

Auditor	Findings	Confirmed by verifier	Fabricated content
ChatGPT	6 major, 6 minor	9 of 9 source-checkable; remainder convergence	none found
Gemini	0 above minor	—	22/26 source quotes; 11/11 document quotes; 4 wrong PMIDs

Danny source-level spot check (replaces Gemini audit) — 2026-10-14

8 of 8 ChatGPT-origin components matched their sources (C042 Treister-Goltzman, C047 Wachtel, C052 Ware 2010b, C055 Gaston, C069/C083 Lucas & Walsh, C072 Levin, C073 Hartman, C074 Mittleman). One refinement (Ware 2010b restfulness vs wakefulness) was verified in the abstract and applied as correction #14. Optional C065/PC8 items not marked.

Whiting 2015 corrected article — read via Danny’s browser, 2026-10-14

Danny’s screenshots (corrected JAMA abstract; JAMA 2016;315(14):1522 “Incorrect Effect Estimate”) confirm:

Ashworth WMD −0.12 (95% CI −0.24 to 0.01; 5 trials).
CINV 47% vs 20%, OR 3.82; pain 37% vs 31%, OR 1.41; NRS pain WMD −0.46 (6 trials).
“an increased risk of short-term AEs with cannabinoids, including serious AEs”.
The notice records corrections on 2015-06-26 (Figure 4), 2015-07-13 (Ashworth value in the abstract) and 2015-11-05 (drug name, Table 1), plus a sleep-sentence fix (nabilone insomnia −3.25; restfulness 0.48).

Still held: the patient-rated NRS spasticity −0.76 and the Table 3 odds ratios. The 2015-08-25 notice page did not render for Danny. The PubMed abstract still shows the uncorrected −0.36, so the PubMed record is stale.

Whiting 2015 Table 3 — Danny’s screenshot, 2026-10-14

Any AE OR 3.03 (2.42–3.80; 29 trials/3,714); serious AE OR 1.41 (1.04–1.92; 34/3,248); withdrawal due to AE OR 2.94 (2.18–3.96; 23/2,755). Matches Gemini’s P2 figures and ChatGPT’s cross-audit. Correction #16. Only the patient-rated NRS spasticity −0.76 remains held for Whiting.

Source file in the research repository: medical/MED01/verification/verification-log.md. Private links, some names and practitioner-record passages are removed from the public copy, and broken character encoding from file conversion is repaired; nothing else is changed.