MED01-P1 — Cross-audit prompts (§M10a)

Status: READY 2026-09-26. Ledger v1.0 (+ correction #1), technical synthesis v1.0, plain-English ledger v1.0 and plain-English synthesis v1.0 are all signed off. Git only.

Rule (§M10a): no single-origin claim is audited solely by its originating model. Gemini audits the ChatGPT-origin claims; ChatGPT audits the Gemini-origin claims; both audit the verifier-found claims, the Both-origin claims (reciprocal), and Claude’s handling throughout. The audit tests Claude’s evidence chain — synthesis sentence → claim → source → extracted result — not the auditor’s own earlier pass. Each audit is filed as secondary-audit-chatgpt.md / secondary-audit-gemini.md, its citations are verified like any other, and both must be filed before P1 LOCK. Escalation: keystone or major-safety disputes the reciprocal audit cannot resolve go to Danny’s source-level review.

Assignments (from ledger v1.0 origin fields)

How to launch (Danny): paste each prompt into that model’s existing project chat (the firewall is lifted; auditors may see everything). Give each auditor the four signed Drive docs by title. Paste the reply back; Claude files it verbatim, verifies its citations, and dispositions each finding.

Prompt for Gemini

You are the cross-auditor for MED01-P1 under PROTOCOL-MEDICAL §M10a. You audit the verifier (Claude), not the researchers. The independence firewall has been lifted: you may read everything in the MED01 Drive folder.

Documents (Google Drive → "Medical Cannabis Knowledge Engine" → "MED01 — Historical Evidence Landscape (1960s–2026)"):
- "MED01 Claim Ledger v1.0 SIGNED OFF (P1, 40 claims) — with correction #1 2026-09-26"
- "MED01-P1 Interim Technical Synthesis v1.0 SIGNED OFF 2026-09-26"
- "MED01-P1 Plain-English Claim Ledger" and "MED01-P1 Plain-English Synthesis" (v0.2 docs = signed v1.0 text)
- "MED01-P1 Source Library additions S-241–S-307"
- "MED01 Verification Log (P1, through ledger v0.2, 2026-09-25)" and "MED01-P1 Critique & Verification v0.1"
- Both raw passes (your own and ChatGPT's) in the raw-inputs subfolders, for provenance only.

YOUR ASSIGNMENT
A. ChatGPT-origin claims — C020, C031, C032. You are the only auditor of these.
B. Verifier-found claims — C001, C015, C017, C020 (the 41-subject datum), C024 (UK 1982). These were surfaced by Claude, not by either researcher, and must be audited by both models.
C. Claude's overturns of your own pass: the Hepler & Frank 1971 magnitude (C017), the PubMed count (C001), the harms/therapeutic balance (C027), the synthesis chronology (C005), the Chang 1979 route framing (C013). Audit whether Claude's ruling is supported by the sources, not whether your original answer was defensible.
D. Shared-origin claims (C002–C014, C016, C018, C019, C021–C023, C026, C030, C034–C036, C038–C040): sample at least ten, including every Accepted efficacy or harm claim.

FOR EACH CLAIM, trace: synthesis sentence → ledger claim → S-source → the source's actual text (open it live; PubMed abstract at minimum). Report:
- Does the source say what the ledger says? (identity + support)
- Is the ledger's disposition (Accepted / Accepted with qualification / Inconclusive) justified by the access level and the evidence class?
- Does any synthesis sentence — technical or plain-English — say more, or express less uncertainty, than its ledger claim? Quote both.
- Drift checks D1–D14: any animal→human, association→cause, symptom→disease, one product→another, one route→all, one dose→general, short→long, surrogate→outcome, one population→broader, within-condition→treatment effect, trial product→retail product, absence→safety?
- For the plain-English versions: does anything read as advice? Does any simplification broaden a diagnosis or population?

RULES
- Point to the source, not to your prior pass. A finding must quote the source text.
- Every citation you use will itself be verified. Give resolvable identifiers; never complete one from memory. In your correction round you supplied five PMIDs that resolved to adjacent papers; do not repeat that.
- Do not research new topics. This is an audit of an existing chain.
- Severity: critical (claim unsupported or wrong) · major (overstated / drift) · minor (wording, metadata) · none.

OUTPUT
1. Trace table: # | synthesis sentence (verbatim, §) | claim ID | S-ID(s) | source supports as worded? | disposition justified? | drift (D#) | plain-English fidelity | finding | severity.
2. Whole-synthesis checks: vote counting by significance? harms given the same rigor as benefits? "how this evidence was found" section honest? conclusions stated as of the cutoff? misleading by omission or ordering?
3. Verifier self-check: for each of Claude's overturns in (C), is the ruling correct, and is the basis recorded?
4. The single most serious problem you found, or an explicit statement that you found none above minor.

Prompt for ChatGPT

You are the cross-auditor for MED01-P1 under PROTOCOL-MEDICAL §M10a. You audit the verifier (Claude), not the researchers. This is distinct from your earlier editorial-calibration passes: those improved the drafts; this tests whether the signed chain holds. 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 Claim Ledger v1.0 SIGNED OFF (P1, 40 claims) — with correction #1 2026-09-26"
- "MED01-P1 Interim Technical Synthesis v1.0 SIGNED OFF 2026-09-26"
- "MED01-P1 Plain-English Claim Ledger" and "MED01-P1 Plain-English Synthesis" (v0.2 docs = signed v1.0 text)
- "MED01-P1 Source Library additions S-241–S-307"
- "MED01 Verification Log (P1, through ledger v0.2, 2026-09-25)" and "MED01-P1 Critique & Verification v0.1"
- Both raw passes in the raw-inputs subfolders, for provenance only.

YOUR ASSIGNMENT
A. Gemini-origin claims — C006 (the CBC and Gill 1970 components), C024 (the Canada component), C028, C029, C033, C037. You are the only auditor of these. Several rest on sources Claude could not open; test whether the Inconclusive/HOLD dispositions are honest and whether any Gemini-origin datum leaked into the synthesis at higher confidence than its ledger disposition allows.
B. Verifier-found claims — C001, C015, C017, C020 (41-subject datum), C024 (UK 1982). Surfaced by Claude; must be audited by both models.
C. The HOLD components — C025 (Congressional Record / NIDA 1976), C028 (IND start and counts; the IOM passage), C029 (54 FR 53767 text), C036 (IOM Executive Summary), C037 (four commissions). Where you can open a primary document Claude could not (you reported reading IOM 1999 on Bookshelf), quote the passage with its location so Claude can record it — but note that your read counts as convergence until Claude reproduces it, and say so in your report.
D. Shared-origin claims (C002–C014, C016, C018, C019, C021–C023, C026, C030, C034–C036, C038–C040): sample at least ten, including every Accepted efficacy or harm claim. Since you reviewed the drafts, deliberately pick claims you did not comment on before.

FOR EACH CLAIM, trace: synthesis sentence → ledger claim → S-source → the source's actual text (open it live). Report:
- Does the source say what the ledger says? (identity + support)
- Is the disposition justified by access level and evidence class?
- Does any synthesis sentence — technical or plain-English — say more, or express less uncertainty, than its ledger claim? Quote both.
- Drift checks D1–D14.
- Plain-English: anything advice-like? any broadened diagnosis or population?

RULES
- Point to the source, not to your prior reviews. A finding must quote the source text.
- Every citation you use will be verified. Resolvable identifiers only; never complete one from memory.
- Do not research new topics. This is an audit of an existing chain.
- Severity: critical · major · minor · none.

OUTPUT
1. Trace table: # | synthesis sentence (verbatim, §) | claim ID | S-ID(s) | source supports as worded? | disposition justified? | drift (D#) | plain-English fidelity | finding | severity.
2. Whole-synthesis checks: vote counting by significance? harms given the same rigor as benefits? "how this evidence was found" section honest? conclusions stated as of the cutoff? misleading by omission or ordering?
3. HOLD report: for each item in (C), what you could open, the quoted passage and location, and whether it supports, contradicts, or does not address the held component.
4. The single most serious problem you found, or an explicit statement that you found none above minor.