MED01-P2 — Critique & Verification (v0.1)

Run: verification run 1 · 2026-10-06 · Claude (verifier) · NCBI E-utilities (ESearch / ESummary / EFetch), Crossref REST, ClinicalTrials.gov API v2, Health Canada DPD API, openFDA drugs@FDA. Passes of record: ChatGPT Stage 1 + Stage 2 (raw-inputs/chatgpt/MED01-P2-chatgpt-stage1-raw.md, …-stage2-raw.md); Gemini Stage 1 rerun + Stage 2 (raw-inputs/gemini/MED01-P2-gemini-stage1-rerun-raw.md, …-stage2-raw.md). Gemini’s superseded first Stage 1 is not verified here (postmortem evidence only). Status: run 1 covers identity (check 1), publication integrity, the canonical counts, and abstract-level claim support (check 2) for the contested and keystone items. Full-text checks, the completeness probe and correction rounds are open (§7). Nothing here is a ledger disposition yet.

1. Independence disclosures (weigh before convergence)

ChatGPT: prior MED01 context visible in the account (memory on); attachments arrived as pasted markdown after a first attempt in which they did not mount; no project file, Drive or Probe Appendix seen in Stage 1 (disclosure filed, …-stage1-disclosure.md).
Gemini: Stage 1 of record is a rerun in a separate Google account with no project-Drive access; isolation statement clean; brief and protocol pasted as text. The superseded first run saw the Probe Appendix header and Q1–Q5 via a Drive search (corrections.md). The rerun was started after that exposure in a different account; residual model-level carry-over cannot be excluded and is disclosed.
Stage 2 for both ran in the same chat as Stage 1, after Stage 1 was filed, as designed.

2. Canonical counts (brief §3A, §4a) — verifier run 2026-10-06

Exact canonical term and URL pattern, datetype=pdat, run live:

Period	Verifier count	ChatGPT	Gemini (rerun)
2000–04	4,335	unreproducible (ESearch blocked; not estimated)	5,618
2005–09	7,451	unreproducible	7,542
2010–14	10,316	unreproducible	9,874
2015–17	8,599	unreproducible	8,365

Total 2000–2017 = 30,701 indexed records (shape: rising through 2010–14; the three-year 2015–17 window has the highest annual rate, ≈2,866/yr vs ≈2,063/yr in 2010–14).

ChatGPT: complied with §4a (no estimate).
Gemini: stated that its container could not reach E-utilities, then reported four counts as “live index search figures” with “±1–2%” variation. None reproduces (+29.6 %, +1.2 %, −4.3 %, −2.7 %). §4a breach (“never smooth, interpolate or estimate a count”); repeats P1 (Gemini’s 11,241, not reproducible). Gemini’s reported query translation (2000/01/01:2004/12/31[Date - Publication]) is plausible in form but was not obtainable from a run it says it could not make.

3. Check 1 — citation identity

3a. ChatGPT — clean

PMIDs: 28 unique in Stage 1, 28 in Stage 2; every PMID resolves to the paper named, with matching DOI where both are given (low title-match scores in the automated pass were all cases where the title was not on the same line; each was checked by hand).
DOIs: all resolve to the named works.
Registrations: NCT00046722, NCT00681538, NCT02091375, NCT01040910 all resolve to the described trials. ISRCTN68314063 (Ware 2010) — ChatGPT itself flagged it as publication-reported, not opened; verifier open pending.
Sativex Canada DIN 02266121: confirmed in the Health Canada Drug Product Database.
Minor: describes NCT00046722 as showing “experimentally induced pain as an outcome”; the current record’s primary outcome is change in HIV neuropathic pain on a 100-mm VAS (matches the paper’s primary). Not an identity error.

3b. Gemini — major identity failure

27 distinct PMIDs are wrong-object or nonexistent across Stage 1 (rerun) and Stage 2, every one marked “Verified live (resolved title copied)”:

Cited as	Gemini PMID → actually resolves to	Correct record
Alger 2002	12445899 → Parri 2002, astrocytes/thalamus	12498988
Cravatt 2001	11752458 → Chin 2001, viperin (CMV)	11470906
Huggins 2012	22442435 → O’Brien 2012, vasospasm/epilepsy surgery	22727500
Collin 2007	17355545 → Zoccolella 2007, riluzole/ALS	17355549
Portenoy 2012	22442249 → Eildermann 2012, primate testis cells	22483680
Zammit 2002	12446536 → Hawton 2002, adolescent self-harm	12446534
Rogeberg 2013	23319617 → Pugach 2013, Holocene gene flow	23319626
Budney 2004	15516668 → Ahlstrom-Emanuelsson 2004, allergic rhinitis	15514394
Allen 2004	15707076 → Bizri 2003, “SARS: the flying death”	15479672
Koppel 2014	24777948 → Vieweger 2015, soil/plant health	24778283
Kerbrat 2016	27806232 → Chokshi 2016, health reform	27806235
Nutt 2013	23764895 → Wang 2013, bile-acid transport	23756634
Grant 2012	22615655 → Saeidnia 2011, genus Achillea	22629287
Rocha 2014	25215638 → Zhang 2014, cervical carcinoma	24142199 (J Neurooncol)
Geffrey 2015	26105080 → does not exist	26114620
Rog 2005	16186535 → Vinci 2005, Charcot-Marie-Tooth QoL	16186518
Van Gaal 2005	15837488 → Diaz 2005, Langmuir-Blodgett films	15836887
Pi-Sunyer 2006	16474120 → Thurlow 2006, HSV packaging protein	16478899
Porter & Jacobson 2013	24238796 → Smith 2013, resting-state fMRI	24237632
Englund 2013	23108272 → Matejko 2013, math scores/white matter	23042808
Stith & Vigil 2016	27330386 → Waheeb 2016, retinopathy of prematurity	27257247
Tashkin 2013	23802294 → Neithercott 2013, “Fathers know best”	23802821
Tzadok 2016	26801975 → Gulland 2016, Ebola vaccine	26800377
Jetly 2015	25467990 → Senchenkov 2015, postmastectomy seromas	25467221
Singh & Bali 2013	24490074 → Catalá 2013, polyunsaturated fatty acids	24474921
Skrabek 2008	18266804 → Molnár 2008, EEG synchronization	17974490
van den Elsen 2015	25969098 → Feigenbaum 2015, laser optics	25972490

Further Gemini identity defects:

DOIs: nonexistent (Crossref 404) — Collin 2007 10.1177/1352458506070613 (true 10.1111/j.1468-1331.2006.01639.x), Cravatt 10.1073/pnas.231193898 (true …161191698), Grant 10.2174/187152712800564491 (true 10.2174/1874205X01206010018), Henquet 10.1136/bmj.38267.431400.7C (true …664086.63), Rogeberg 10.1073/pnas.1220329110 (true …1215678110), Wade 2004 10.1177/135245850401000408 (true 10.1191/1352458504ms1082oa), Singh & Bali 10.12659/PCR.889317 (true 10.1159/000356446). Wrong-object: Geffrey 10.1111/epi.13076 (→ Lal 2015, RBFOX1 deletions; true 10.1111/epi.13060), Huggins 10.1016/j.pain.2012.02.030 (→ Taylor 2012, rTMS; true …04.020), Rocha 10.1590/0100-3984.2013.1909 (→ Candido 2015, sclerosing peritonitis; true 10.1007/s11060-013-1277-1).
Sativex DINs 02269926 and 02268571 (two different values in the two Gemini runs) do not exist in the Health Canada database; the product’s DIN is 02266121.
Author misattribution of the two bibliometric anchors: “Chaiton et al., 2022” is Ng & Chang 2022; “Sweileh et al., 2018” is Matielo et al. 2018 (Crossref).
Titles marked “resolved title copied” that are not the resolved titles: Wilson & Nicoll 2001 (“…at physiological synapses”; actual “…hippocampal synapses”), Huggins 2012 (truncated and reworded), Portenoy 2012 (an invented subtitle, “a randomized, addition-on, controlled trial”), Englund 2013 (reworded), Porter & Jacobson 2013 (“use” dropped). Rule 11 (quote by copy) and rule 10 (resolved title copied) both fail here.
Access level: every Gemini reference is marked “Full text”, including paywalled items; not credible as a blanket statement and not used to lift any §M9 cap.

Net: Gemini’s substantive paper choices are mostly real and relevant, and every one of the 27 intended papers was located by the verifier — but the identifiers attached to them were not opened. This is the P1 pattern (wrong-object PMIDs) at roughly four times the volume, in a pass that labels each one verified live.

3c. Publication integrity (§M1) — dated 2026-10-06

No PubMed publication-type flag (Retracted Publication / Expression of Concern) and no Crossref updated-by record on any resolved journal identifier in either pass. Correction 2026-10-13: this check did not read PubMed’s ErratumIn links. A rescan of all 127 P2 PMIDs found errata or updates on Whiting 2015 (four errata, including “Incorrect Values Reported”, 2015-08-25, and “Incorrect Effect Estimate”, 2016-04-12), Devinsky 2016, Bachhuber 2014, Van Gaal 2005, Pi-Sunyer 2006 and Wilson & Nicoll 2001, and a 2019 Cochrane update of Marshall 2014. The Whiting errata change a ledgered figure (C078); the others are held for content checks. Koppel 2014 carries [RETIRED] in its PubMed title (AAN guideline-retirement status, flagged by ChatGPT); usable as a period source with the status stated.

3d. Regulatory records checked

Syndros, NDA 205525: original approval 2016-07-01 (openFDA) — both passes correct.
Cesamet, NDA 018677: original approval 1985-12-26; the 2006-05-15 entry is a labeling supplement (SUPPL11), not an approval. Gemini’s “FDA approved Cesamet, May 15, 2006” is not supported; ChatGPT’s refusal to count the 2006 relaunch as a P2 approval is correct.

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

Item	Finding	Who
CAMS (Zajicek 2003)	Abstract: no effect on Ashworth (p=0.40); extract vs placebo 0.32 (−1.04 to 1.67); THC vs placebo 0.94 (−0.44 to 2.31); patient-reported spasticity/pain improved (p=0.003); funder per PubMed MS Society	ChatGPT exact. Gemini’s THC CI “−0.36 to 2.24, p=0.14” not supported; “funded by the UK MRC” not supported; “up to 78% unblinding” not in abstract (full-text check)
Ware 2010	9.4% vs 0%: difference 0.7, 95% CI 0.02–1.4; drug inhaled TID for the first five days of each 14-day cycle; 23 recruited / 21 completed	ChatGPT exact. Gemini’s “−0.7, CI −1.37 to −0.03, p=0.023” and “three times daily for 14-day periods” not supported
Abrams 2007	Parallel-group RCT (“randomly assigned to smoke either cannabis or placebo”); 34% vs 17% median reduction, p=0.03; >30% responders 52% vs 24%, p=0.04	ChatGPT exact. Gemini “crossover” and “p=0.016” not supported
Novotna 2011	572 enrolled, 272 responders, 241 randomized; ITT difference P=0.0002	Both correct; Gemini’s “−0.84 (−1.29 to −0.40)” not in abstract (full-text check)
Portenoy 2012	Primary 30%-responder analysis not significant (P=.59); secondary continuous responder analysis favoured low (P=.008) and medium (P=.039) doses	Gemini CAND-23 reports only the secondary result as “statistically significant pain reduction” — drift D11 / primary-outcome omission. ChatGPT did not make the claim
Rog 2005	66 patients; 4-week treatment; −2.7 vs −1.4	Gemini supported
Johnson 2010	177 randomized (60/58/59); primary favoured THC:CBD vs placebo	Gemini supported
Zammit 2002	50,087; adjusted OR for >50 uses 6.7 (2.1–21.7) in the cannabis-only group	Gemini supported
Moore 2007	any psychotic outcome, ever-use: pooled adjusted OR 1.41 (1.20–1.65)	Both supported
Meier 2012	1,037; decline concentrated in adolescent-onset users; cessation did not fully restore functioning	Both supported; Gemini’s “8 IQ points” not in abstract (full-text check)
Whiting 2015	79 trials/6,462; 4 low RoB; CINV OR 3.82 (1.55–9.42); pain OR 1.41 (0.99–2.00); NRS pain WMD −0.46; Ashworth WMD −0.36 (−0.69 to −0.05; 7 trials); moderate-quality for chronic pain and spasticity	Gemini’s spasticity figure “WMD −0.76 (−1.38 to −0.14; 3 trials), though objective spasticity scales showed no significant effect” is contradicted by the abstract; its AE ORs 3.03 / 2.94 need full text
Asbridge 2012	9 studies; OR 1.92 (1.35–2.73); fatal 2.10 (1.31–3.36)	Gemini supported
Allen 2004	19 identified, 9 presented; cessation ended vomiting in 7; abnormal washing in 9 of 10	Both broadly supported
Pletcher 2012	cohort of 5,115; nonlinear association, FEV1/FVC slightly higher at low exposure	Gemini’s “N=5,016” not in abstract (possibly analytic sample — full-text check)
Koppel 2014	34 studies, 8 Class I; MS spasticity: OCE effective, nabiximols/THC probably effective for patient-centred measures; “possible both OCE and THC are effective for … objective measures at 1 year”	ChatGPT supported. Gemini describes Koppel as an AAN guideline on “complementary and alternative medicine in MS” with “Level A … established as ineffective” for Ashworth — scope and wording do not match Koppel; likely conflation with a different AAN guideline (unverified)
Kerbrat 2016	6 on active 50 mg/day; 4 consented to report; syndrome in 3 of 4; one brain dead	ChatGPT supported. Gemini’s “five healthy volunteers … cerebral hemorrhages and necrosis” (S1) and “one death and four hospitalizations” (CAND-14) are not in the abstract and are inconsistent with each other
Naftali 2013	Primary (remission, CDAI<150) not achieved; response 10/11 vs placebo	ChatGPT supported. Gemini’s “without endoscopic mucosal healing” is not in this study — matches Naftali 2021 (oral CBD; “clinical but not endoscopic remission”): anachronism (P3 source read into P2). Verifier-found: current NCT01040910 primary outcome is “reduction of CDAI by 70 points” (response), not remission — registered-vs-published primary mismatch to check against the registry history
Geffrey 2015	13 on clobazam; clobazam +60%, N-desmethylclobazam +500% at 4 weeks; doses reduced in 10/13	ChatGPT supported. Gemini’s “three- to five-fold … driven by CYP2C19 inhibition” — mechanism not in abstract
Tramèr 2001	30 randomised comparisons, 1,366 patients	ChatGPT exact. Gemini’s “30 randomized controlled trials” matches the full text (“We analysed data from 30 randomised controlled trials”); the earlier “comparisons ≠ trials” remark is withdrawn (cross-audit 2026-10-13)
Devinsky 2016	214 / 162 / 137; AEs 79%; SAEs 30%; median motor-seizure reduction 36.5%	Both supported (ChatGPT correctly labels it uncontrolled)
Gaston 2017	changes in topiramate, rufinamide, N-desmethylclobazam; AST/ALT higher with valproate	ChatGPT supported
Lucas & Walsh 2017	271 respondents; 63% substitution; 30% opioids	ChatGPT supported
Hazekamp 2013	5,540 patients, 2003–2010	ChatGPT supported

5. Drift and rule findings

#	Finding	Researcher	Rule
1	Four canonical counts reported despite stated inability to run the query; none reproduces	Gemini	§4a
2	27 wrong-object/nonexistent PMIDs, 10 bad DOIs, 2 nonexistent DINs, all labelled “verified live”	Gemini	rules 10, 10a; §M1 check 1
3	“Resolved title copied” titles that are not the resolved titles	Gemini	rules 10, 11
4	Portenoy: secondary result reported as the finding; failed primary omitted	Gemini	D11; rule 4
5	Naftali “no endoscopic healing” taken from a 2021 study	Gemini	rule 6a (hindsight)
6	Koppel scope/rating language misattributed	Gemini	check 2
7	Whiting spasticity estimate contradicted by the abstract	Gemini	check 2
8	Cesamet 2006 labeling supplement presented as an approval	Gemini	§M8
9	Several numerical details differ from abstracts (CAMS CI/funder, Ware CI/duration, Abrams design/p, Kerbrat counts)	Gemini	check 2
10	ISRCTN68314063 not opened (self-flagged)	ChatGPT	rule 10a — verifier to open
11	Mücke 2018 Cochrane used in Gemini’s era-table certainty column although retrospective	Gemini	rule 6a (labelled retrospective in the reference list but used as a rating)

ChatGPT run 1: no identity failure, no unsupported number found in the items checked; it consistently labelled uncontrolled change as uncontrolled (Devinsky 2016, Abrams 2011) and kept retrospective sources separate.

6. Convergence map (for the ledger) — provisional

Weight convergence after §1: ChatGPT memory-on; Gemini identifiers unreliable (topic convergence counts, identifier convergence does not).

Both, Stage 1: CAMS · Novotna · Abrams 2007 · Ware 2010 · Devinsky 2017 · Tramèr 2001 · Whiting 2015 · NASEM 2017 · Moore 2007 · Meier 2012 · Kerbrat/van Esbroeck (BIA 10-2474) · rimonabant EU authorization/suspension · Sativex Canada 2005 & 2007 · Sativex UK 2010 · Ng & Chang 2022 · Portenoy 2012 (ChatGPT as survey line only).
Both, Stage 2: Syndros 2016 · Guzmán 2006 · Porter & Jacobson 2013 · Devinsky 2016 · Russo 2004/2011/2016 · Bachhuber 2014 (+ Shover 2019 retrospective) · US 6,630,507 (granted 2003-10-07) · ACP 2008 · Naftali 2013 · Skrabek 2008 · Tomida 2006 · AAP 2015 · CMCR.
ChatGPT only: Budney 2003 · Gunn 2016 · Hartman 2015 · Mittleman 2001 · Gaston 2017 · Blake 2006 · Gloss 2014 (Cochrane epilepsy) · Koppel 2014 (correctly characterised) · Wachtel 2002 · Ilan 2005 · Bhattacharyya 2010 · TGA AusPAR (Sativex) · AEMPS Sativex record · Levin 2011 · Lucas & Walsh 2017 · Reiman 2017 · Treister-Goltzman 2018/19 (retrospective) · Hazekamp 2013 · US 9,474,726 · ACOG · co-twin cognition 2017.
Gemini only: Wilson & Nicoll 2001 · Laprairie 2015 · Cravatt 2001 · Huggins 2012 · Rog 2005 · Johnson 2010 · Nurmikko 2007 · Wallace 2015 · Wilsey 2008/2013 · Ellis 2009 · MUSEC 2012 · Collin 2007/2010 · Wade 2004 · Zammit 2002 · Arseneault 2002 · Henquet 2005 · Asbridge 2012 · Allen 2004 (CHS) · Pletcher 2012 · Rogeberg 2013 · Di Forti 2015 · Volkow 2014 · Craker v. DEA · AMA 2009 · Nutt 2013 · Stith & Vigil 2016 · Englund 2013 · Hazekamp 2006 · Abrams 2007b · Boehnke 2016 · Jetly 2015 · Bergamaschi 2011 · van den Elsen 2015 · Tzadok 2016 · Rocha 2014 · Singh & Bali 2013.
Public-claim labels differ: PC5 (ChatGPT unsupported, Gemini partly supported); PC8 (ChatGPT partly supported, Gemini contradicted). Agree on PC1 contradicted, PC2 partly, PC3 unsupported, PC4 contradicted, PC6 partly, PC7 contradicted.
Stage-2 “omissions” lists (ChatGPT 12, Gemini 6) are postmortem evidence on the two-stage design: both researchers missed the probe topics in Stage 1 (patent grant, CECD, Bachhuber, professional bodies, epilepsy development pathway).

6a. Completeness probe (§M12b) — verifier, 2026-10-06

Logged PubMed queries (T = (cannabinoid*[ti] OR cannabis[ti] OR marijuana[ti] OR marihuana[ti] OR nabiximols[ti] OR sativex[ti] OR dronabinol[ti] OR nabilone[ti] OR cannabidiol[ti] OR tetrahydrocannabinol[ti])), compared against every PMID in the four passes of record plus the verifier-corrected Gemini PMIDs:

Query	Hits	Not cited by either pass
`T AND (systematic review[pt] OR meta-analysis[pt] OR "systematic review"[ti] OR "meta-analysis"[ti]) AND 2000:2017[dp]`	227	219
`T AND "Cochrane Database Syst Rev"[ta] AND 2000:2017[dp]`	16	15
`T AND randomized controlled trial[pt] AND 2000:2017[dp] AND humans[mh]`	547	519

Most uncited hits are cannabis-use-disorder treatment, prevention or epidemiology studies (in scope only for harms). Screened by title for medical-cannabinoid trials in patients (70 records) and for indication/harm reviews (110 records). Items judged material to the P2 map — origin verifier-found, mandatory §M10a cross-audit targets:

Keystone omissions (trials)

Zajicek 2013, CUPID (Lancet Neurol; PMID 23856559) — dronabinol vs placebo on progression of progressive MS; the only large human disease-modification RCT identified in the window; reported no effect. Directly bears on both passes’ absence statements about disease modification.
Cannabis-In-Cachexia Study Group (Strasser) 2006 (J Clin Oncol; PMID 16849753) — cannabis extract vs THC vs placebo in cancer anorexia-cachexia; a period head-to-head extract-vs-THC comparison (probe Q5) as well as a cachexia trial.
Jatoi 2002 (J Clin Oncol; PMID 11786587) — dronabinol vs megestrol vs combination in cancer anorexia (Gemini’s era table alludes to it without a citation).
Meiri 2007 (PMID 17355735) — dronabinol vs ondansetron for delayed CINV (the rare P2 comparison with a modern antiemetic).
Frank 2008 (BMJ; PMID 18182416) — nabilone vs dihydrocodeine, chronic neuropathic pain.
Svendsen 2004 (BMJ; PMID 15258006), Berman 2004 (PMID 15561385), Selvarajah 2010 (PMID 19808912), Kavia 2010 (PMID 20829244), Corey-Bloom 2012 (smoked cannabis, MS spasticity; PMID 22586334), Collin 2010 (PMID 20307378), Vaney 2004 (PMID 15327040), CAMS 12-month follow-up (PMID 16291891).
Leweke 2012 (CBD vs amisulpride in schizophrenia; PMID 22832859), Müller-Vahl 2002/2003 (THC in Tourette; PMIDs 11951146, 12716250), Abrams 2003 (short-term safety in HIV; PMID 12965981), Naftali 2017 (low-dose CBD in Crohn’s, negative; PMID 28349233).

Keystone omissions (syntheses)

Wang 2008, “Adverse effects of medical cannabinoids: a systematic review” (CMAJ; PMID 18559804) — the principal period harms synthesis for therapeutic cannabinoids; neither pass cited it.
Cochrane reviews 2000–2017 not cited: Smith 2015 CINV (26561338; used in P1 as retrospective, a period source here), Lutge 2013 HIV/AIDS (23633327), Krishnan 2009 dementia (19370677), Walitt 2016 fibromyalgia (27428009 — ChatGPT referred to it without an identifier), Curtis 2009 Tourette (19821373), McLoughlin 2014 cannabis and schizophrenia (25314586), Marshall 2014 pharmacotherapies for cannabis dependence (25515775).
Pain syntheses: Campbell 2001 (11440935; P1 retrospective use), Iskedjian 2007, Martín-Sánchez 2009, Lynch 2011 & 2015, Andreae 2015 (inhaled cannabis IPD meta-analysis), Deshpande 2015, Nugent 2017 (Ann Intern Med; pain + general harms).
MS: Wade 2010 meta-analysis of nabiximols for spasticity (20558502); Lakhan 2009.
Blinding: Wright 2012 (Trials; PMID 23046749) — assessment of blinding in Sativex studies; directly relevant to §3E.
Harms: Marconi 2016 dose–response psychosis (26884547); Large 2011 earlier onset of psychosis (21300939); Lev-Ran 2014 depression (23795762); Mehra 2006 lung cancer (16832000); Calabria 2010 mortality (20565525); Li 2012 crashes (21976636); Conner 2016 neonatal outcomes (27607879); Grant 2003 and Schreiner 2012 residual neurocognitive effects (12901774, 22731735); Degenhardt 2013 global burden (24204649).

Boundary check: the Lennox-Gastaut CBD trials (Thiele, Lancet, online January 2018; Devinsky, NEJM, May 2018) are P3; Gemini handled Thiele correctly as a boundary case. Records with 2018 PubMed dates in the SR set (e.g. Häuser 2018, Schussel 2018) are retrospective.

6b. Registry cross-check (§M4) — verifier, 2026-10-09 (disclosure control, not a publication-bias analysis)

Method. ClinicalTrials.gov API v2, interventional studies with completion date 2000-01-01 → 2017-12-31, intervention queries: Sativex/nabiximols · dronabinol/Marinol/Syndros · nabilone/Cesamet · cannabidiol/Epidiolex/GWP42003 · smoked/vaporized cannabis. 329 unique records; 290 confirmed by intervention name; 246 COMPLETED. Of these, 112 had no posted results and no linked results reference; 31 of those are found in PubMed by NCT number ([si]); 19 PubMed lookups errored and need a rerun; 62 have no trace. After removing healthy-volunteer, pharmacokinetic and cannabis-use-disorder studies from the 62 no-trace records, 32 patient/therapeutic trials remain (rechecked 2026-10-10: none of the 32 has a PubMed record by NCT number; the 19 errored records are not yet screened).

Second pass, title/sponsor/year matching (PubMed does not index every NCT number). Candidate matches, checked 2026-10-09 against PI, site, product, design and N in the paper (✅ = confirmed; ✗ = rejected):

✅ NCT00272207 → Skrabek 2008 (17974490)
✅ NCT00381199 → Ware 2010, nabilone sleep (20007734)
✅ NCT00760695 → Andries 2014 (24105610)
✅ NCT00623376 → Pooyania 2010 (20434606)
✅ NCT00959218 (Bionorica, n=240) → Schimrigk 2017 (29073592). Primary endpoint not met (NRS −1.92 vs −1.81, p=0.676); the abstract’s conclusion (“a safe long-term treatment option”) rests on the safety extension — a negative-primary trial published eight years after completion
✅ NCT00480181 → Turcotte 2015 (25288189)
✅ NCT00248378 → Corey-Bloom 2012 (22586334)
✗ NCT00309413 → Leweke 2012 (22832859) — rejected: the Leweke 2012 record lists NCT00439634. NCT00309413 moves to the no-publication list
✅ NCT00153192 (Brigham and Women’s) → Narang 2008 (18088560)
✅ NCT00079560 (NIDA) → Haney 2007 (17589370)

No publication identified in the documented searches (possible non-reporting; to be disclosed in the synthesis after one more title search each):

NCT00309413 — (moved from candidate list 2026-10-09)
NCT01898520 — Jazz/GW, Sativex for spasticity in children with cerebral palsy, phase 3, n=72, completed 2017-03
NCT00642499 — Solvay, dronabinol for HAART-associated symptoms, phase 2, n=103, completed 2005
NCT00123201 — Solvay, dronabinol metered-dose inhaler for migraine, phase 2, completed 2007
NCT00699634 — nabilone for phantom-limb pain, phase 2, n=50, completed 2011
NCT00380965 and NCT00381264 — NEMA Research, Cesamet in neuropathic and MS pain, phase 4, n=23 each, completed 2007
NCT01538225 — Almirall, neurophysiological Sativex study in MS spasticity, n=45, completed 2013
NCT00254761 and NCT00241579 — CMCR, smoked cannabis in neuropathic and experimental pain (title search pending; may correspond to Wallace 2007 or Ellis 2009)
NCT02051387 — Mannheim, CBD drug-interaction study in schizophrenia, completed 2017-08 (results may be post-2017; boundary)

Not run: WHO ICTRP and EU-CTR (no API reached in this run). EU-registered GW/Almirall nabiximols trials are therefore not covered; disclose as a limitation.

6d. Registry re-screen — verifier, 2026-10-13 (after cross-audit finding B2-02)

The trigger was NCT01898520, listed as having no publication identified. It was published in 2020 (Fairhurst C et al., Dev Med Child Neurol; PMID 32342496). The title search scheduled in §6b had never been run.

Method.

All 132 completed records with no posted results and no results reference were re-queried in PubMed by NCT number, without the earlier intervention-name filter: 44 hits, 88 none, 0 errors.
The 88 were screened by an explicit rule: a cannabinoid given to patients for a therapeutic purpose.
The 28 that qualified were searched by title, investigator, sponsor, product and design.

Excluded as non-therapeutic. These 12 were among the original “32 patient trials”:

NCT01025700, NCT00683488, NCT01853020, NCT02567344, NCT00783185, NCT00142961, NCT02723149, NCT00624715 (dependence treatment, behavioural or challenge studies);
NCT02112292, NCT01740960, NCT00241579, NCT01893424 (healthy-volunteer or pharmacokinetic studies).

Also excluded: NCT01311778, which matches Manini 2015, a healthy-volunteer pharmacokinetic study (PMID 25748562).

Matched (17).

The 9 from §6b: NCT00272207, NCT00381199, NCT00760695, NCT00623376, NCT00959218, NCT00480181, NCT00248378, NCT00153192, NCT00079560.
NCT00254761 → Wilsey 2008 (18403272)
NCT00781001 → Wallace 2015 (25843054)
NCT00238550 → Selvarajah 2010 (19808912)
NCT00255580 → Ellis 2009 (18688212)
NCT00642512 → Meiri 2007 (17355735)
NCT01318369 → de Vries 2016 (26505163; THC vs active placebo diazepam, n = 24)
NCT00812851 → Weber 2010 (20498181; THC for cramps in ALS)
NCT01898520 → Fairhurst 2020 (32342496; retrospective)

No publication identified (11).

NCT00642499 (Solvay, dronabinol in HAART nausea, n = 103)
NCT00123201 (Solvay, dronabinol MDI in migraine)
NCT00380913, NCT00380965, NCT00381264 (NEMA, Cesamet × 3)
NCT00699634 (nabilone, phantom limb, n = 50)
NCT01222468 (nabilone in spinal-cord injury, n = 12, 2015)
NCT01538225 (Almirall, Sativex neurophysiology in MS, n = 45)
NCT00309413 (Cologne, CBD crossover in schizophrenia, n = 29)
NCT01983267 (Rambam, cognition in cancer patients using cannabis, n = 100)
NCT02051387 (Mannheim, CBD interactions, completed 2017-08; boundary)

The search results are saved with the round files. This section supersedes the §6b counts (112 / 31 / 62 / 19 / 32 / 9).

6c. Official-record checks — verifier, 2026-10-09

Record	Finding	Bears on
NASEM 2017, The Health Effects of Cannabis and Cannabinoids (NAP reader, Summary, ch. 2; DOI 10.17226/24625)	Opened (DOC). Conclusive/substantial evidence of therapeutic effect: chronic pain in adults (cannabis, 4-1); CINV (oral cannabinoids, 4-3); patient-reported MS spasticity (oral cannabinoids, 4-7a). Moderate: short-term sleep outcomes (4-19). Limited: HIV/AIDS appetite/weight (4-4a), clinician-measured MS spasticity (4-7a), Tourette (4-8), social anxiety (CBD, 4-17), PTSD (nabilone, 4-20). Limited evidence of no effect: dementia (4-13), glaucoma IOP (4-14), depressive symptoms (4-18). No or insufficient: cancers incl. glioma (4-2), cancer cachexia / anorexia nervosa (4-4b), IBS (4-5), epilepsy (4-6). Harms: substantial association with schizophrenia/other psychoses, highest risk among most frequent users (12-1); substantial for motor-vehicle crashes (9-3); maternal smoking and lower birth weight (10-2); respiratory symptoms and chronic bronchitis episodes with long-term smoking (7-3a); moderate for paediatric overdose injuries in states where legal (9-4b); moderate evidence of no association with lung cancer from smoking (5-1); moderate for acute impairment of learning, memory and attention (11-1a). Recommendation 4: regulatory and supply barriers to research.	Both passes cite NASEM; their summaries are consistent with these conclusion numbers. Epilepsy rating is “insufficient” in NASEM’s cut-off (Aug 2016), before the 2017 NEJM Dravet trial — a dated-synthesis point, not a contradiction
Health Canada — Sativex (DIN 02266121; NOC database)	NOC with conditions 2005-04-15 (control 091289; MS neuropathic pain, adjunctive); NOC/c 2007-08-01 (control 109461; advanced cancer pain, adjunctive). Manufacturer GW Pharma Ltd. Both conditional indications were removed 2019-12-12 (confirmatory studies did not confirm benefit) — retrospective, rule 6a	Both passes correct on dates; neither mentioned the 2019 removal (outside window; record as later reinterpretation)
GOV.UK / MHRA — Sativex	UK marketing authorisation “in June 2010” for MS spasticity; Home Office named-patient import licence statement for 2006. Exact day (16 vs 21 June) not verifiable from the official text opened	ChatGPT and Gemini disagree on the day; month confirmed
EMA — Acomplia (rimonabant), EMEA/H/C/000666	Marketing authorisation 2006-06-19; withdrawn 2009-01-16 (MA holder request after CHMP recommended suspension, Oct 2008, psychiatric risk)	Both passes correct on the event sequence
openFDA — Epidiolex, NDA 210365	Original approval 2018-06-25	Outside P2; both passes correctly treat it as P3. FDA orphan-designation dates for CBD in Dravet/LGS: OOPD query returned no usable output — not verified
US 6,630,507 B1 (Google Patents)	Priority 1998-04-21; filed 1999-04-21; granted 2003-10-07; assignee US DHHS; inventors Hampson, Axelrod, Grimaldi; expired 2019-04-21; title “Cannabinoids as antioxidants and neuroprotectants”	ChatGPT exact; Gemini’s grant date matches
US 9,474,726 B2	Priority 2014-06-17; granted 2016-10-25; original assignee GW Pharma; claim 1 = CBD in the treatment of FIRES (febrile infection-related epilepsy syndrome)	ChatGPT’s “including Dravet syndrome” mischaracterises claim 1 — new finding against ChatGPT (D11-type scope error, not identity)
US 9,066,920 B2	GW Pharma + Otsuka; priority 2009-07-03; granted 2015-06-30; claim 1 = CBD ≥400 mg/day for partial seizures	Gemini’s “Dravet/LGS” scope not supported
US 8,071,641 B2	Assignees Hadasit / Yissum (inventors include Mechoulam, Gallily); “Treating or preventing diabetes with cannabidiol”; granted 2011-12-06	Gemini’s attribution to GW/THC:CBD formulations not supported
Craker v. DEA, 714 F.3d 17 (1st Cir.)	Decided 2013-04-15, No. 09-1220; petition for review of the DEA’s denial of a cannabis-manufacturer registration — denied	Gemini correct
81 FR 53846 (DEA)	2016-08-12, doc. 2016-17955: policy statement on applications to become registered to manufacture marijuana for research (ends single-supplier arrangement in principle)	Both correct
82 FR 14815 (DEA)	2017-03-23: Syndros (dronabinol oral solution) placed in Schedule II	ChatGPT correct

7. Open before ledger v1.0

~~Completeness probe~~ (§6a) · ~~registry cross-check~~ (§6b; 9/10 matches confirmed) · ~~official records~~ (§6c). Still open: rerun 19 errored PubMed-by-NCT lookups; ICTRP/EU-CTR not covered (limitation); one more title search for each no-publication trial.
Full text not read (claims capped at abstract level): CAMS unblinding % and capsule composition; Novotna adjusted difference; Whiting AE/withdrawal ORs; Meier “8 IQ points”; Geffrey CYP2C19 mechanism; TGA AusPAR and AEMPS records (ChatGPT only); MHRA exact day; NCT01040910 registry history; FDA orphan-designation dates.
~~Correction rounds~~ — done 2026-10-06 (ChatGPT 3/3; Gemini 38 of 55 confirmed, counts admitted fabricated, 4 new failures).
~~Researcher-performance record~~ — in the verification log.

Source file in the research repository: medical/MED01/critique-P2.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.