codex-math

Codex Math

Use Codex as an adversarial mathematical co-processor. The goal is not to sound convincing; the goal is to separate proved facts, real gaps, false alarms, and promising leads.

Treat every generated argument as a lead, not a verdict. A proof that looks polished can still be false. A reviewer objection can still be a false positive. Triage every finding.

Mode Selection

• Verify when the user provides a proof, theorem, lemma, or paper section and wants an audit.
• Write when the user provides a precise result and asks for a proof.
• Explore when the result may be false, under-specified, or hard enough that counterexample search and sufficient conditions are valuable.

For high-stakes claims, use two passes: one constructive pass and one hostile pass. In Codex environments that expose subagents, spawn independent verifier/explorer agents only when the user explicitly asks for parallel or independent agents.

Verify Mode

1. Extract the exact statement, hypotheses, definitions, and proof block.
2. Build a dependency list: equations, lemmas, citations, domain restrictions, and hidden regularity assumptions.
3. Walk the proof step by step. Classify each step as:
   • OK: justified by a definition, prior result, explicit algebra, or standard theorem.
   • GAP: a necessary implication is missing.
   • ERROR: the step is false, with a counterexample or algebraic correction.
   • UNCLEAR: the step might be right but needs a missing definition, citation, or convention.
4. Check signs, dimensions, boundary cases, quantifiers, and whether examples are being used as proof.
5. Triage findings. Report only issues you can explain concretely.

Verification output:

## Verdict
PASS | PASS WITH CAVEATS | FAIL | INCONCLUSIVE

## Findings
1. Severity: blocking | major | minor
   Location:
   Issue:
   Why it matters:
   Suggested fix:

## Checked
- Definitions:
- Algebra/signs:
- Edge cases:
- Citations or external results:

Write Mode