MED01-P2 — Cross-audit prompts (§M10a)

Status: READY 2026-10-12. Signed and ready for audit: ledger P2 v1.0 (+ correction #4), interim technical synthesis P2 v1.0, plain-English ledger P2 v1.0 and plain-English summary P2 v1.0. Each audit is filed verbatim as secondary-audit-P2-chatgpt.md / secondary-audit-P2-gemini.md. Its citations and quotations are verified like any other, and both audits must be filed before P2 LOCK.

Rule (§M10a): the audit tests Claude’s evidence chain (synthesis sentence → ledger claim → S-source → the source’s actual text), not the auditor’s own earlier pass. No single-origin component is audited only by the model that proposed it:

• Gemini audits ChatGPT-origin components.
• ChatGPT audits Gemini-origin components.
• Both audit the verifier-found components, the held items and a sample of Both-origin claims.
• Each auditor also audits Claude’s rulings against the other researcher, which gives an independent check of every overturn.

Escalation: keystone or major-safety disputes the reciprocal audit cannot resolve go to Danny’s source-level review.

Lessons applied from P1 and earlier P2 steps:

• Three batches per auditor. The first §M11 attempt stopped without reviewing anything because the job was too big for one reply.
• Quotes by copy only, each with its location. Gemini’s P1 cross-audit contained six “verbatim” quotations that were not in the document.
• Identifiers opened, never completed from memory. Gemini’s P2 passes had 27 wrong-object PMIDs, all labelled “verified live”.

Assignments (from the origin fields of ledger P2 v1.0)

How to launch (Danny):

• ChatGPT: a new chat in the project account.
• Gemini: a chat in the Google account that has access to the project Drive (the main account, not the second account used for the Stage 1 rerun, which cannot see the Drive). The independence firewall is lifted, so the auditors may read everything.
• In each chat, paste the main prompt (Batch 1), then the Batch 2 and Batch 3 follow-ups as each reply comes back.
• Paste all three replies here. Claude files them together, verbatim, verifies the auditor’s own citations and quotations, and dispositions each finding.

Drive documents both auditors read (folder “Medical Cannabis Knowledge Engine” → “MED01 — Historical Evidence Landscape (1960s–2026)”):

• “MED01 Ledger — P2 v1.0 SIGNED OFF 2026-10-10 (with correction #4)”
• “MED01-P2 Interim Synthesis v1.0 SIGNED OFF 2026-10-12”
• “MED01-P2 Plain-English Claim Ledger v1.0 SIGNED OFF 2026-10-12”
• “MED01-P2 Plain-English Summary v1.0 SIGNED OFF 2026-10-12”
• “MED01-P2 Critique & Verification v0.1 (2026-10-10, C067 denominator note)”
• “MED01 Verification Log (through P2 record checks, 2026-10-09)”
• “MED01-P2 Source Library additions v1.0 (S-308–S-445)” (the v0.1 Doc had six wrong Year cells, so use v1.0; identical rows are in knowledge-core “SOURCE_LIBRARY.part3of4.md (rev. 2026-10-10 year fix)”)
• The P2 raw passes in the raw-inputs subfolders, for provenance only.

Prompt for Gemini

You are the cross-auditor for MED01 Part 2 (medical cannabis and cannabinoid research, 2000–2017) under PROTOCOL-MEDICAL §M10a. You audit the verifier (Claude), not the researchers. The independence firewall is lifted: you may read everything in the MED01 Drive folder.

DOCUMENTS (Google Drive → "Medical Cannabis Knowledge Engine" → "MED01 — Historical Evidence Landscape (1960s–2026)"):
- "MED01 Ledger — P2 v1.0 SIGNED OFF 2026-10-10 (with correction #4)" — claims MED01-C041 to C085; the authority
- "MED01-P2 Interim Synthesis v1.0 SIGNED OFF 2026-10-12"
- "MED01-P2 Plain-English Claim Ledger v1.0 SIGNED OFF 2026-10-12" and "MED01-P2 Plain-English Summary v1.0 SIGNED OFF 2026-10-12"
- "MED01-P2 Critique & Verification v0.1 (2026-10-10, C067 denominator note)" and "MED01 Verification Log (through P2 record checks, 2026-10-09)"
- "MED01-P2 Source Library additions v1.0 (S-308–S-445)" — full citations, PMIDs and DOIs for every S-ID
- The P2 raw passes in the raw-inputs subfolders, for provenance only.
If a document is missing or will not open, say which. That is the only reason to stop.

THIS AUDIT RUNS IN THREE BATCHES, one per message. Do Batch 1 now, then wait.

BATCH 1 — ChatGPT-origin components. You are the only auditor of these:
C042 (Treister-Goltzman) · C047 (Wachtel, Ilan, Bhattacharyya, Russo 2011) · C052 (Ware 2010 nabilone, Blake 2006) · C055 (Gloss, Gaston) · C063 (the 2017 Schedule II rule; Cesamet 2006 not counted as an approval) · C065 (US 9,474,726) · C069 (Hazekamp 2013, Lucas & Walsh) · C072 (Budney 2003, Levin) · C073 (Hartman) · C074 (Tashkin 2012, Mittleman) · C075 (Gunn, ACOG) · C081 (Gloss, Walitt) · C082 (ACOG, AAP, ACP) · C083 (Lucas & Walsh, Reiman, Abrams 2011).
Also audit Claude's rulings against ChatGPT: C065 (US 9,474,726 claim 1 is FIRES, not Dravet) and C085 PC8 ("contradicted as worded"; ChatGPT had "partly supported"). Is each ruling supported by the source?

BATCH 2 (next message) — verifier-found. Full audit of the keystones C041, C054, C060, C067, C077 (Wang 2008, Nugent 2017) and C084, plus a sample of at least 10 verifier-found components from the list in the ledger's "Mandatory §M10a cross-audit targets" line and the origin fields.

BATCH 3 (last message) — a sample of at least 10 Both-origin claims (C043, C046, C048 CAMS/Novotna, C050 Abrams/Ware, C053, C055 Devinsky 2016/2017, C057, C058, C062, C066, C078, C079, C085), including every Accepted efficacy or harm claim among them; then the whole-document checks and your summary.

FOR EACH CLAIM OR COMPONENT, trace: synthesis sentence → ledger claim → S-source → the source's actual text. Open the source live: the PubMed abstract at minimum, or the official record. Report:
- Identity: does the S-ID resolve to the paper or record named?
- Support: does the source say what the ledger says, with the same numbers, population, product and outcome status?
- Disposition: is the ledger's disposition (Accepted / Accepted with qualification / Inconclusive) justified by the access level (FT / AB / DOC / BIB) and the evidence class?
- Downstream fidelity: does any sentence in the technical synthesis or either plain-English document say more, or express less uncertainty, than the claim? Quote both.
- Drift D1–D14: animal→human, association→cause, symptom→disease, one product→another, one route→all, one dose→general, short→long, surrogate→outcome, one population→broader, secondary→primary, trial product→retail product, absence→safety, hindsight (2018+ read into the period).

RULES
- QUOTE BY COPY ONLY. Every quotation, from a source or from a MED01 document, must be copied from the page you opened, with its location (PubMed abstract, section number, row ID). Claude will diff every quotation against the original. In the P1 cross-audit, six "verbatim" quotations from you were not in the document; that must not recur. If you cannot copy it, paraphrase and label it "paraphrase".
- IDENTIFIERS: give only identifiers you opened in this session. Never complete a PMID or DOI from memory. In P2 you supplied 27 PubMed IDs that pointed to unrelated papers; any new identifier you give will be resolved by the verifier.
- Do not research new topics. This audits an existing chain.
- A complete batch with no findings is a valid result.
- Severity: critical (claim unsupported or wrong) · major (overstated, drift, wrong disposition) · minor (wording, metadata) · none.

OUTPUT, EACH BATCH
Trace table: # | claim ID + component | S-ID(s) | identity ok? | source supports as worded? (quote + location) | disposition justified? | drift (D#) | downstream fidelity (synthesis / plain-English) | finding | severity.
At the end of each batch: "Batch N complete. Last item checked: …".

BATCH 3 ONLY, AFTER THE TABLE
1. Whole-document checks: vote counting by significance? harms given the same rigor as benefits? Is the "how this evidence was found" section honest? Are conclusions stated as of 2017 with later events labelled? Anything misleading by omission or ordering?
2. Verifier self-check: for each of Claude's rulings in Batch 1, correct or not, and is the basis recorded?
3. The single most serious problem you found, or an explicit statement that you found none above minor.

Gemini follow-ups (same chat):

Batch 2: the verifier-found keystones C041, C054, C060, C067, C077 (Wang 2008, Nugent 2017) and C084 in full, plus a sample of at least 10 verifier-found components. Same trace method, rules and table. Continue the numbering.

Batch 3: the Both-origin sample (at least 10: C043, C046, C048 CAMS/Novotna, C050 Abrams/Ware, C053, C055 Devinsky 2016/2017, C057, C058, C062, C066, C078, C079, C085), then the whole-document checks, the verifier self-check and the single most serious problem. Same rules; continue the numbering.

Prompt for ChatGPT

You are the cross-auditor for MED01 Part 2 (medical cannabis and cannabinoid research, 2000–2017) under PROTOCOL-MEDICAL §M10a. You audit the verifier (Claude), not the researchers. This is distinct from your earlier calibration, editorial and fidelity reviews: those checked that the documents matched the ledger; this tests whether the ledger matches the sources. The firewall is lifted: read everything in the MED01 Drive folder.

DOCUMENTS (Google Drive → "Medical Cannabis Knowledge Engine" → "MED01 — Historical Evidence Landscape (1960s–2026)"):
- "MED01 Ledger — P2 v1.0 SIGNED OFF 2026-10-10 (with correction #4)" — claims MED01-C041 to C085; the authority
- "MED01-P2 Interim Synthesis v1.0 SIGNED OFF 2026-10-12"
- "MED01-P2 Plain-English Claim Ledger v1.0 SIGNED OFF 2026-10-12" and "MED01-P2 Plain-English Summary v1.0 SIGNED OFF 2026-10-12"
- "MED01-P2 Critique & Verification v0.1 (2026-10-10, C067 denominator note)" and "MED01 Verification Log (through P2 record checks, 2026-10-09)"
- "MED01-P2 Source Library additions v1.0 (S-308–S-445)" — full citations, PMIDs and DOIs for every S-ID
- The P2 raw passes in the raw-inputs subfolders, for provenance only.
If a document is missing or will not open, say which. That is the only reason to stop.

THIS AUDIT RUNS IN THREE BATCHES, one per message. Do Batch 1 now, then wait.

BATCH 1 — Gemini-origin components. You are the only auditor of these. Most had wrong identifiers in Gemini's pass that Claude replaced; check that Claude's replacement is the paper Gemini meant AND that the ledger's figures come from that paper:
C044 · C045 · C047 (Englund) · C048 (Wade 2004, Collin 2007/2010, MUSEC) · C049 (Rog) · C050 (Ellis, Wilsey, Wallace) · C051 (Johnson, Nurmikko) · C055 (Tzadok) · C056 (Bergamaschi, Jetly, van den Elsen) · C059 (Singh & Bali, Rocha) · C061 (Hazekamp 2006, Abrams 2007b) · C064 (RIO trials, Huggins) · C065 (US 9,066,920, US 8,071,641) · C068 (Craker v. DEA, Nutt, Stith & Vigil, Grant) · C069 (COMPASS) · C070 (Zammit, Arseneault, Henquet) · C071 (Rogeberg) · C072 (Budney 2004, Volkow) · C073 (Asbridge) · C074 (Pletcher, Tashkin 2013) · C076 (Allen).
Also audit Claude's rulings against Gemini. Is each one supported by the source?
- C041: Gemini's counts withdrawn as fabricated
- C048: CAMS THC interval and funder
- C050: Abrams design and p-value; Ware CI and dosing days
- C051: Portenoy primary outcome
- C057: Naftali "no endoscopic healing", attributed to a 2021 study
- C063: Cesamet 2006 = labeling supplement
- C064: Kerbrat counts
- C065: US 9,066,920 and US 8,071,641 scope
- C078: Whiting spasticity estimate
- C080: Koppel scope
- C085: PC5 "unsupported" (Gemini had "partly supported")

BATCH 2 (next message) — verifier-found. Full audit of the keystones C041, C054, C060, C067, C077 (Wang 2008, Nugent 2017) and C084, plus a sample of at least 10 verifier-found components, chosen from the ledger's "Mandatory §M10a cross-audit targets" line and the origin fields. Also the 10 HELD items at the end of ledger P2: for each, say whether you could open the full text or primary document Claude could not. If you can, quote the passage with its location. Your read counts as convergence until Claude reproduces it; say so.

BATCH 3 (last message) — a sample of at least 10 Both-origin claims (C043, C046, C048 CAMS/Novotna, C050 Abrams/Ware, C053, C055 Devinsky 2016/2017, C057, C058, C062, C066, C078, C079, C085), including every Accepted efficacy or harm claim among them. Then the whole-document checks and your summary. You researched many of these, so deliberately check the ones you did not comment on before.

FOR EACH CLAIM OR COMPONENT, trace: synthesis sentence → ledger claim → S-source → the source's actual text. Open the source live: the PubMed abstract at minimum, or the official record. Report:
- Identity: does the S-ID resolve to the paper or record named?
- Support: does the source say what the ledger says, with the same numbers, population, product and outcome status?
- Disposition: is it justified by access level and evidence class?
- Downstream fidelity: does any synthesis or plain-English sentence say more, or express less uncertainty, than the claim? Quote both.
- Drift D1–D14, including hindsight (2018+ read into the period).

RULES
- Quote by copy only, with location; Claude will diff every quotation. Paraphrases are labelled "paraphrase".
- Identifiers only if opened in this session; never complete one from memory.
- Do not research new topics. This audits an existing chain.
- A complete batch with no findings is a valid result.
- Severity: critical · major · minor · none.

OUTPUT, EACH BATCH
Trace table: # | claim ID + component | S-ID(s) | identity ok? | source supports as worded? (quote + location) | disposition justified? | drift (D#) | downstream fidelity | finding | severity.
At the end of each batch: "Batch N complete. Last item checked: …".

BATCH 2 ALSO
HELD report: for each held item, what you could open, the quoted passage and location, and whether it supports, contradicts or does not address the held component.

BATCH 3 ONLY, AFTER THE TABLE
1. Whole-document checks: vote counting by significance? harms given the same rigor as benefits? Is the "how this evidence was found" section honest? Are conclusions stated as of 2017 with later events labelled? Anything misleading by omission or ordering?
2. Verifier self-check: for each of Claude's rulings in Batch 1, correct or not, and is the basis recorded?
3. The single most serious problem you found, or an explicit statement that you found none above minor.

ChatGPT follow-ups (same chat):

Batch 2: the verifier-found keystones C041, C054, C060, C067, C077 (Wang 2008, Nugent 2017) and C084 in full, a sample of at least 10 verifier-found components, and the HELD report for the 10 held items. Same trace method, rules and table. Continue the numbering.

Batch 3: the Both-origin sample (at least 10: C043, C046, C048 CAMS/Novotna, C050 Abrams/Ware, C053, C055 Devinsky 2016/2017, C057, C058, C062, C066, C078, C079, C085), then the whole-document checks, the verifier self-check and the single most serious problem. Same rules; continue the numbering.