adversarial-proof-audit

name	adversarial-proof-audit
description	Independently attack a candidate mathematical proof: find gaps, circular lemmas, silent hypothesis changes, insufficient uniformity, and false completions. Use after any draft proof of an open problem, before claiming success, or when the user asks to audit, stress-test, or red-team a proof.

Adversarial proof audit

Every candidate argument must be challenged by an independent adversarial pass. In the source workflow this is continuous, not a single end ceremony.

Independence

Prefer a different model or a fresh subagent that did not author the draft. Do not defend the draft; try to break it.

Audit checklist

For the draft proof and each cited lemma:

1. Claim match — Does the proved statement equal the problem’s principal question, or a weaker substitute banned by the prompt?
2. Hidden strengthenings — Extra assumptions (smoothness, primes only, almost all (n), subsequence, asymptotic regime only)?
3. Circular reductions — Is a key lemma equivalent in strength to the goal?
4. Uniformity / ranges — Are error terms strong enough for the claimed main term? Are theorems applied inside their hypotheses?
5. Definition drift — Did minimization targets, quantifiers, or discrete vs continuous relaxations change mid-proof?
6. Computational leaps — Are numerics presented as proofs? Are finite checks converted into certificates correctly?
7. Case leaks — Missing edge cases, empty configurations, small-(n) gaps?
8. Citation accuracy — Named theorems stated correctly?
9. Adversarial counterexamples — Attempt to falsify intermediate lemmas with constructions or computation.
10. Blocked-route smell — Does the “proof” stop at “it suffices to show [open statement]”?

Output format (AUDIT.md)

# Audit — <problem id> — <date>

## Verdict
PASS | FAIL | PASS_WITH_REPAIRS_REQUIRED

## Claim match
...

## Findings
### F1 — <title> — severity: blocker|major|minor
- Location:
- Issue:
- Suggested repair or counterexample:

## Counterexample / computation attempts
...

## Required repairs before claiming success
- [ ] ...

Policy

• FAIL if any blocker remains.
• Authors must repair or abandon routes; do not paper over blockers with “routine” or “clearly.”
• Re-run audit after material repairs until PASS.