MED01 — Secondary Audit (cross-audit template)
Run after Claude’s verification, ledger, and synthesis. Governed by PROTOCOL-MEDICAL §M10a: no single-origin claim is audited solely by its originating model; shared-origin (Both) claims get reciprocal audit, with origin-independent escalation for keystones, major safety claims, and unresolved disputes. Each auditor’s work is filed as its own copy of this template — secondary-audit-chatgpt.md, secondary-audit-gemini.md — and registered in AUDIT_LOG.md. The round cannot lock until both are filed.
| Claim origin | Audited by |
|---|---|
ChatGPT |
Gemini |
Gemini |
ChatGPT |
verifier-found, or originated / overturned by Claude |
both |
Both |
reciprocal audit by both |
What is audited: Claude’s final evidence chain and handling — not the auditor’s own earlier pass. A finding must point to the source, not to what the auditor originally wrote.
Trace path: synthesis sentence → MED01 claim → shared S-source → extracted result (extraction block)
Auditor: ___ · Date: ___ · Audited versions: synthesis v___ · ledger v___ · PROTOCOL-MEDICAL v___
Claims assigned to this auditor (by origin): ___ · Claims excluded as own-origin: ___
Sampling within the assignment: every Accepted efficacy and harm claim, every keystone, every verifier-found claim, and every claim where Claude overturned a researcher; others sampled — state how.
Trace table
| # | Synthesis sentence (verbatim, § ref) | Claim ID | Origin | S-ID(s) | Evidence base (type · design per source) | Comparative estimate is design-appropriate? (§M4) | Does the source support the sentence as worded? | Drift found (D1–D14) | Wording stronger than ledger? | Finding | Severity |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 |
Severity: critical (unsupported or wrong) · major (overstated / drift) · minor (wording, metadata) · none
Whole-synthesis checks
| Check | Finding |
|---|---|
| Vote counting by significance anywhere? Is the weighting of conflicting studies stated and reasoned? (§M6) | |
| Harms given the same rigor as benefits; “no evidence” not worded as safety? (§M4, D14) | |
| “How this evidence was found, and what may be missing” section present and honest; no systematic-review claim? (§M12b) | |
| Conclusions stated as of the search cutoff? (§M12a) | |
| Misleading by omission, emphasis, or ordering even though each claim is correct? |
Plain-English fidelity check
Does each plain-English statement keep the technical version’s qualifiers, evidence strata (source type and study design), and population limits? Is there a “Who this does not tell us about” line wherever the finding is bounded (§M11)? Does anything read as advice (§M0)?
| Plain-English statement | Technical counterpart | Uncertainty preserved? | Indication broadened? | Advice-like? | Finding |
|---|---|---|---|---|---|
Auditor’s own citations
Any source or methodological standard the auditor cites is verified by Claude like any other (identity + support) before the finding is acted on.
| Cited by auditor | Check 1 | Check 2 |
|---|---|---|
Resolution
| Finding # | Action taken | Correction log # | Escalated to origin-independent check? (§M10a) | Status (resolved / documented-open) |
|---|---|---|---|---|