MED01-P2 — §M10a cross-audit — ChatGPT — RAW, verbatim

Received 2026-10-13 by Danny’s paste: three batch replies to cross-audit-prompts-P2.md, from one new ChatGPT chat. Formatting notes: ChatGPT’s reply began by echoing the prompt, which is shortened here to a pointer. Table cells that the paste split across lines are joined into single rows, and the citation chips the paste carried (e.g. “PubMed +1”) are kept in square brackets. The wording is unchanged. Verifier checks are in secondary-audit-P2-verification.md and the dispositions in secondary-audit-P2-dispositions.md.

[Echoed prompt: identical to the ChatGPT prompt in cross-audit-prompts-P2.md.]

Worked for 8m 12s

MED01-P2 §M10a Cross-Audit — Batch 1 Scope: Gemini-origin components and Claude’s rulings against Gemini, using the signed P2 ledger with correction #4 as the authority.

All seven required signed and verification documents were accessible. I also retrieved the source library and Gemini’s verification history.

Confirmed finding C048 — Wade 2004: Minor outcome-description error. The signed ledger describes Wade’s nonsignificant result as a “primary composite.” However, the paper identifies its primary outcome as:

“a Visual Analogue Scale (VAS) score for each patient’s most troublesome symptom.”

— Wade et al. 2004, PubMed abstract, outcome measures. [pubmed.ncbi.nlm.nih.gov]

This was a patient-specific symptom measure, not a composite score. Claude correctly preserved the negative primary result and the positive spasticity-specific result, but the endpoint label needs correction. The synthesis and plain-English documents do not appear to repeat that incorrect label.

Trace table — Gemini-origin components Reading the table: “Identity” means the corrected S-ID identifies the intended source, not that Gemini’s original identifier was correct. AB means PubMed abstract, DOC an official or primary record, BIB bibliographic identification only. All unquoted audit descriptions are paraphrases. The downstream column quotes short passages from the signed technical synthesis (T) and plain-English documents (P); repeated wording is cited once when several components share a claim.

C074 correction required The signed ledger currently says:

“low cumulative exposure was not associated with lower lung function (FEV1/FVC slightly higher)”

The full text of Pletcher distinguishes the separate measures and notes:

“a positive association for both FEV1 and FVC at low levels of exposure”

— Pletcher et al., Discussion. [PMC]

The expression FEV₁/FVC normally denotes the ratio, an important indicator of airflow obstruction. It is not interchangeable with “FEV₁ and FVC.”

Recommended ledger correction: “At low cumulative exposure, FEV₁ and FVC were slightly higher; associations were nonlinear, and the study did not establish protection against pulmonary disease.”

The source’s full text confirms 5,115 recruited and 5,016 included in the main complete-data analysis, so Claude’s correction of the analytical denominator stands. [PMC]

Trace table — Claude’s disputed rulings against Gemini These are judgments of Claude’s verification decisions, not new dispositions of the researchers’ original passes.

Batch 1 disposition Critical 0 · Major 1 · Minor 2 · Rulings upheld 12/12 The 12/12 refers to the listed Claude rulings, not independent reproduction of every underlying search or extraction. C041 remains limited by the fact that I did not rerun the canonical PubMed query.

Across the checked components, Claude’s replacement identities generally lead to the papers Gemini intended, and the ledger usually preserves the correct numbers, population, product, endpoint and evidence class. More importantly, the signed chain has retained negative primary outcomes, uncontrolled-design limitations and observational-versus-causal distinctions.

I recommend a logged correction to C074 before treating its pulmonary-function detail as verified, plus the C048 primary-outcome label correction and the optional C065 Type II specificity correction. Neither synthesis nor the plain-English pair appears to reproduce the two outcome-label errors.

The remaining access limits are explicit: S-312 and S-407 are bibliographic-only; S-425 has no PubMed abstract; and the C041 count calculation was not independently reproduced. Those limits are not silently promoted to full-text verification.

Batch 1 complete. Last item checked: C085 — PC5, Claude’s “unsupported” ruling.

I’ll stop here for Batch 2.

[Danny: Batch 2 follow-up prompt, as in cross-audit-prompts-P2.md]

Worked for 8m 12s

MED01-P2 §M10a Cross-Audit — Batch 2 Audit date: October 13, 2026 Authority: P2 signed ledger v1.0 with correction #4 Scope: Verifier-found keystones, additional verifier-found components, and all ten HELD groups.

All seven required documents remain accessible. The numbering continues from Batch 1, beginning at #44.

A. Keystone trace table Two important distinctions have emerged: the primary clinical trial results generally match the signed ledger, but one harms statistic has lost its original denominator. The registry-derived counts also require more caution than an independently reproduced database query.

Finding B2-01 — Wang denominator transfer Severity: Major · C077 · S-435 · Source-level confirmed

Claude’s reported percentages and rate ratios are accurate, but the claim does not clearly restrict the 96.6% denominator to the randomized trials. The technical synthesis repeats the ambiguity.

The actual review comprised 23 randomized trials and eight observational studies. The 4,779 adverse events and the 96.6% calculation belong to the randomized-trial evidence. [PMC]

Suggested correction:

Wang et al. reviewed 31 studies (23 randomized trials and eight observational studies). Within the 23 randomized trials, 4,615 of 4,779 adverse events reported in cannabinoid-treated participants (96.6%) were classified as non-serious. Non-serious events occurred more frequently than in controls (RR 1.86, 95% CI 1.57–2.21); serious-event rates did not differ significantly (RR 1.04, 95% CI 0.78–1.39). Long-term safety remained inadequately characterized.

This correction matters because the number of adverse events is not the number of patients experiencing an adverse event. The original statistic must not be read as a 96.6% probability of a non-serious event.

Finding B2-02 — A published trial is still counted among the unresolved registry records Severity: Major · C067 · S-403 · Source-level confirmed

The ledger and plain-English summary name the 72-child nabiximols trial, NCT01898520, as an example for which no publication was identified. But its results are traceable in two primary records:

The EU Clinical Trials Register published results for the corresponding study on October 4, 2018. [EU Clinical Trials Register]

Fairhurst et al. published the randomized trial on April 27, 2020, explicitly identifying NCT01898520 and the 72 randomized participants. [Wiley Online Library]

The published abstract states:

“There was no significant difference in the spasticity 0 to 10 NRS”

— Fairhurst et al., 2020, Abstract, Results. [Wiley Online Library]

The trial also reported three hallucination cases, including one associated with a suicide attempt. This is important negative efficacy and safety evidence, although both results were published after the P2 historical window. [Wiley Online Library]

This is not evidence that the study had been published by December 31, 2017. It is evidence that Claude’s 2026 verification search missed an available publication and an EU results record.

Required disposition: Remove NCT01898520 from the untraced-publication examples, document the 2018 and 2020 retrospective findings, and rerun the matching process before retaining the count of 23 unresolved trials. The corrected count cannot simply be assumed to be 22 without verifying the entire candidate set and search rules.

Finding B2-03 — Kerbrat: Claude rejected a count that the primary regulatory record supports Severity: Major · C064 HELD item · S-394 and French IGAS report

Claude was correct that the 2016 NEJM article reports detailed clinical findings for four consenting participants, three with serious neurological illness. But the broader regulatory investigation reports one volunteer who died and four additional volunteers hospitalized with severe neurological disorders, plus a sixth exposed volunteer hospitalized for monitoring without symptoms.

The French government’s May 2016 IGAS report states:

“Quatre autres volontaires du même groupe de l’essai présentant des troubles neurologiques graves”

— IGAS Report 2016-012R, p. 13, §1 (four other volunteers in the same trial group presenting severe neurological disorders). [igas.gouv.fr]

The primary records therefore describe different denominators, not necessarily contradictory counts. Claude’s narrow ledger wording remains defensible because it explicitly refers to the four participants in Kerbrat’s report. However, the verification ruling should not treat Gemini’s “one death and four hospitalizations” as an unsupported underlying event count. It is supported by the contemporaneous regulatory record, albeit not by the four-person publication subset. [New England Journal of Medicine +1]

This HELD component can now be marked convergent on the regulatory cohort counts, subject to Claude reproducing the IGAS passage. The report’s text was accessible, although its PDF page-image rendering failed.

B. Remaining keystone traces C054 — HIV and anorexia components; C067 — trial reporting; C084 — bounded absences

C054 keystone result: The ledger preserves the fundamental contrast: the large cancer-cachexia trials were negative for cannabinoids relative to their comparators, while HIV calorie-intake findings came from small and different populations. The Cochrane conclusion prevents the laboratory results from being promoted to established clinical efficacy.

C060 keystone result: CUPID’s negative primary result and three-year follow-up were faithfully carried forward. “No effect on progression” should remain shorthand for no statistically demonstrated treatment difference in this trial, not proof that cannabinoids cannot modify any disease.

C067 keystone result: The disclosure framing is sound, but the numerical reconstruction is not yet independently repeatable, and the missed 2020 publication is a substantive verification failure. This remains the central open keystone.

C084 keystone result: The bounded-absence wording is appropriate. Its registry-publication component inherits the C067 correction requirement.

C. Additional verifier-found sample The sample below covers 11 additional verifier-found components from the mandatory target areas. It deliberately includes negative trials, a meta-analysis, a Cochrane review, a harm synthesis, and a blinding assessment.

Sample finding: No additional correction above minor emerged from these 11 components. Claude’s strongest verification behavior here is retaining the correct primary outcome status even when an author’s abstract conclusion is optimistic.

D. HELD report — all ten groups The status column distinguishes resolved, partly resolved, and still held. “Convergence” means that I opened a primary record or full text that Claude did not document opening; Claude must reproduce the specific passage before the verification chain is closed.

Finding B2-04 — Whiting’s serious-event signal should be made visible downstream Severity: Major · C077/C078 · S-402 · Source-level confirmed

Whiting’s 2015 systematic review found statistically higher odds of serious adverse events, not merely more minor short-term side effects:

Source: Whiting et al., JAMA 2015, Table 3. [JAMA Network]

The existing synthesis says:

“Non-serious events were more frequent … serious events were not”

—that is correctly attributed to Wang 2008, but the later Whiting serious-event estimate is not equally visible in the therapeutic-harms discussion. The plain-English summary similarly emphasizes non-serious side effects.

Suggested addition, dated and attributed: “In a separate 2015 systematic review covering a different, larger set of randomized trials, Whiting et al. reported increased odds of serious adverse events (OR 1.41, 95% CI 1.04–1.92), although study quality and short follow-up limited interpretation.”

This is not evidence that Wang’s finding was erroneous. The two meta-analyses used different included studies and analytical measures; they should not be compared as if they were repeated measurements of one population.

E. Batch 2 disposition Critical 0 · Major findings 4 · HELD groups reviewed 10/10 · Extra components 11 Four major findings requiring Claude’s disposition:

B2-01 — Wang denominator: 96.6% is a proportion of recorded events from the 23 randomized trials, not all 31 included studies.

B2-02 — C067 publication miss: NCT01898520 has published trial results and must no longer appear as an unresolved publication without qualification.

B2-03 — Kerbrat ruling: The larger French regulatory cohort supports one death and four other symptomatic hospitalizations; distinguish it from the four-person NEJM subset.

B2-04 — Whiting harms: The serious-adverse-event odds ratio is verifiable in full text and deserves appropriately qualified disclosure alongside the 2008 harms synthesis.

The C041 canonical counts remain unreplicated, and the C067 aggregate registry denominator remains incompletely auditable without the underlying record-level export. These are access and reproducibility limitations, not independently established errors in those counts.

Most remaining verifier-found clinical components passed source-level checks. The stronger conclusion from this batch is that Claude generally maintained trial-outcome accuracy, but a signed evidence chain can still lose denominator precision, miss publication records, or apply too narrow an interpretation of a verifier ruling.

Batch 2 complete. Last item checked: C082 — ACP/AMA professional-body recommendation content (HELD group 10).

[Danny: Batch 3 follow-up prompt, as in cross-audit-prompts-P2.md]

Worked for 2m 57s

MED01-P2 §M10a Cross-Audit — Batch 3 Audit date: October 13, 2026 Scope: Both-origin claims, whole-document checks, and verifier self-audit Numbering: Continues from Batch 2 at #68

Major finding: C078 contains a superseded numerical estimate Major — correction required I found a source-level problem that changes the earlier audit’s assessment of Claude.

The signed ledger uses Whiting et al. (2015) to report an Ashworth spasticity weighted mean difference of −0.36 (95% CI −0.69 to −0.05; seven trials).

That figure appears in the PubMed abstract. However, JAMA issued an official correction in August 2015, replacing it with:

“weighted mean difference, −0.12 (95% CI, −0.24 to 0.01); 5 trials.”

— JAMA, “Incorrect Values Reported,” published August 25, 2015. [JAMA Network]

The publisher’s corrected full article also distinguishes the two outcome measures:

Both figures are explicitly reported in the corrected JAMA article, under “Spasticity Due to MS or Paraplegia.” [JAMA Network]

This means Claude was correct to reject Gemini’s attribution of −0.76 to Ashworth, but incorrect to replace it with an uncorrected Ashworth estimate. The distinction is clinically important because the corrected findings reinforce the ledger’s broader conclusion that patient-reported and clinician-rated spasticity outcomes did not consistently agree.

Required correction: Update C078 and the C048/C066 interpretation notes where applicable, preserve both outcome-specific estimates, and add the official JAMA correction to S-402’s citation-integrity record. Claude’s previous ruling should be marked partly correct, not simply upheld.

1. Batch 3 trace table — Both-origin claims Every required claim is included. Components within multi-source claims are separated when they bear independently on efficacy, harm, design or regulatory interpretation.

Notation: AB = abstract opened; FT = full text opened; DOC = official record opened. All commentary outside quotation marks is paraphrase. T = signed technical synthesis; P = signed plain-English ledger or summary.

Batch 3 findings requiring correction B3-01 — Whiting corrected Ashworth effect (major). This is the source-version failure documented at the beginning of Batch 3. It changes the clinician-measured spasticity result from statistically significant to nonsignificant. The official 2015 correction is decisive.

B3-02 — Devinsky population drift (major). The plain-English ledger says “214 children,” whereas the study’s eligibility was ages 1–30. Replace this with “214 children and young adults with childhood-onset treatment-resistant epilepsy.” The corrected wording should not imply every enrolled individual was a minor.

B3-03 — Meiri equivalence wording (minor). The 61-person trial had observed total-response rates of 54% with dronabinol and 58% with ondansetron. Those similar point estimates do not demonstrate equivalence or noninferiority. Suggested wording: “The small trial reported similar response rates, but it was not designed to establish that the treatments were equivalent.” [PubMed]

B3-04 — Plain-English summary cross-references (minor). Its concluding line says:

“no advantage for cannabinoids in cancer weight loss [C048, C050, C053, C054]”

Only C054 directly supports the large cancer-cachexia trials. C048 concerns MS spasticity, C050 neuropathic pain, and C053 chemotherapy nausea. Remove the extraneous cross-references. The substantive claim is supported, but the trace chain is defective.

B3-05 — Tramèr verifier wording (minor). Claude treated Gemini’s “30 randomized trials” as erroneous because the abstract says “30 randomised comparisons.” However, the review’s full text itself says:

“We analysed data from 30 randomised controlled trials”

— Tramèr et al., Results, Studies. [PMC]

The distinction is still worth retaining for precision, but describing Gemini’s wording as a source contradiction is not justified. The original article uses both formulations.

2. Whole-document checks I checked the signed technical ledger, technical synthesis, plain-English ledger and plain-English summary against the source-level findings accumulated across the three batches.

Vote counting and research quality The synthesis uses a generally defensible evidence-map structure. It does not turn a collection of statistically significant individual studies into a blanket claim of treatment efficacy.

For example, it places the negative CAMS clinician-rated result beside patient-reported benefits and the responder-enriched Novotna trial, while also acknowledging that enrichment restricts applicability.

The corrected Whiting analysis strengthens this distinction: the publisher reports a nonsignificant Ashworth estimate but a statistically significant patient-rated NRS estimate. [JAMA Network]

A useful improvement would be to give this contrast explicitly in the technical synthesis rather than relying on the generalized phrase “clinician-rated and patient-reported outcomes did not consistently agree.”

Harms and benefits The research system did not neglect harms. It includes psychosis associations, withdrawal, dependence, driving impairment, respiratory findings, adverse drug interactions, serious trial adverse events, and special-population risks.

The problem is not a wholesale pro-benefit bias. It is that the verification precision applied to clinical effects was not consistently maintained for the harms syntheses.

In particular, a reader could encounter Wang 2008’s nonsignificant serious-event difference without seeing Whiting 2015’s higher serious-event odds. Both are legitimate results from different reviews, and both should be given their appropriate place and limitations. [JAMA Network]

Search transparency The technical synthesis correctly calls itself an evidence map, not a systematic efficacy review. Its plain-English summary explains that sources were found through database searches, a targeted completeness probe and selected official-record checks.

That transparency is a strength. However, the absolute formulation that every figure was checked against its source should be narrowed to acknowledge that abstract-level checking did not always detect publisher corrections.

A suitable addition would state that verification was performed at the documented access level, with corrections, inaccessible full texts and later source updates managed through the verification log.

Hindsight I did not find a new case of unlabelled post-2017 scientific findings being imported into the signed P2 therapeutic conclusions.

The 2019 reversal of the opioid-law association is clearly retrospective. The 2019 removal of Canada’s conditional Sativex indications is also correctly separated from the initial 2005/2007 events. NASEM’s 2017 epilepsy rating is explicitly contextualized as preceding the Dravet RCT. [PMC +2]

The 2020 publication uncovered for NCT01898520 should likewise be recorded as a later publication of a trial completed in 2017, not as evidence published by the end of the P2 window.

3. Verifier self-check — Claude’s Batch 1 rulings This table revisits the rulings against Gemini, rather than the correctness of the final ledger claims alone.

Correct 9 · Partly correct 2 · Supported with access limit 1 The verdicts above supersede my provisional 12/12 ruling in Batch 2.

What this reveals about the verification process Claude did well at recovering real sources from incorrect identifiers, checking trial populations and numerical endpoints, preserving failed primary outcomes, and refusing to invent information where access was unavailable.

The newly discovered failure mode is narrower and more important: verifying the identity of a paper and matching its PubMed abstract is not always enough to verify its final published numerical result.

In Whiting’s case, the paper identity was correct, the PMID was correct, and the PubMed abstract contained the number Claude copied. Nonetheless, the publisher had already formally corrected that result within the historical period.

This is why the corrected journal article and its notices should take precedence over stale abstract text.

4. Single most serious problem across all three batches The unrecognized 2015 publisher correction to Whiting’s clinician-rated spasticity estimate (C078, S-402).

I rank this as the most serious because it meets four conditions simultaneously:

It affects an Accepted claim in a major systematic review.

The numerical correction changes whether the Ashworth estimate is statistically significant.

The corrected source existed before the 2017 evidence-map cutoff.

Claude’s verification process explicitly rejected Gemini’s alternative estimate while overlooking the publisher’s correction history.

The registry-publication miss in C067 is also major and requires correction. But it does not directly change an Accepted treatment-effect estimate inside a period synthesis. C078 does.

Required source-chain correction The repaired C078 evidence should distinguish:

Ashworth, clinician-rated: WMD −0.12, 95% CI −0.24 to 0.01; five trials; nonsignificant.

NRS, patient-rated: mean difference −0.76, 95% CI −1.38 to −0.14; three trials; statistically significant.

Source status: Whiting 2015, original publication followed by a formal 2015 JAMA correction. [JAMA Network +1]

Do not replace C078’s historical evidence-rating statements with present-day efficacy claims. The correction concerns what the 2015 review actually reported after correction, not a new treatment verdict.

5. Final disposition and cross-batch summary Critical 0 · Major 6 · Minor 6 · Unresolved access/provenance gaps 2 Unique findings consolidated across the three batches, without double-counting the same error when it appears in multiple documents. The two highlighted provenance gaps concern the C041 canonical counts and the unreproduced C067 registry denominator chain. The six major problems are: C074’s FEV₁/FVC interpretation, C077’s Wang event denominator, C067’s missed publication, C064’s overly broad Kerbrat count rejection, C078’s corrected Whiting estimate, and C055’s downstream age-population drift. The underemphasis of Whiting’s serious-adverse-event result is a linked harm-reporting issue requiring an addition, rather than a separate independent major count.

Recommended sign-off state: corrections required before declaring §M10a closed. This does not invalidate the entire Part 2 evidence map. Most sampled source identities and trial conclusions held up, but the signed chain needs explicit corrections rather than silent amendments.

For Claude, the immediate sequence should be to update the source library and critique/verification log first, amend C078 and the other affected claims second, reconcile the technical synthesis third, and only then revise the plain-English derivatives. The remaining HELD items should keep their access statuses until the primary-record passages are reproduced.

Batch 3 complete. Last item checked: C085 — PC8, CBD’s psychoactive effects and risks.

Three-batch cross-audit complete.