proof-aware-crypto-tooling-agent/AGENTS.md

AGENTS.md

Guidance for future Codex runs in this repository:

• Do not add Charon or Aeneas transpilation flows.
• Do not regenerate Lean from Rust. Treat shipped Lean files in verified repos as the artifact.
• Do not claim end-to-end verification unless theorem coverage actually exists.
• Keep macOS and Apple Silicon portability. Do not rely on GNU timeout, free, taskset, Docker, Nix, or x86_64 assumptions.
• Prefer explicit theorem-boundary language over marketing language.
• Update tests with every behavior change.
• Do not silently lower risk ratings. A lower score must explain the failed or missing evidence.
• Do not hide proof failures behind warnings.
• Keep claim cards machine-readable and reports explicit about proven claims, preconditions, exclusions, trusted base, and residual risk.
• Consequence-producing commands must be policy gated. Do not build wallet or trading-agent artifacts from R3 arithmetic evidence.
• Distinguish verifier capability failures from proof failures. Missing Mathlib, missing Aeneas, or a missing pinned env script means local replay is unavailable; do not treat it as a clean proof.
• Third-party proof-checking attestations are allowed only as an explicit trust transformation. They must identify the provider, subject repo/commit, theorem names, observed axioms, and signature status. Untrusted attestations must not authorize builds.
• Transparency receipts must be verified against the exact attestation bytes, Signed Tree Head, log public key, and inclusion proof before they can authorize consequences.
• Keep the Merkle log RFC 9162-style unless a new standard is deliberately adopted and documented. Do not replace it with an ad hoc hash chain.
• Do not pretend ML-DSA exists. If no real ML-DSA backend is available, record the signature slot as unavailable and fail closed for policies that require both Ed25519 and ML-DSA.
• Provider private keys and transparency log state belong under ignored provider/state/ or provider/out/ paths. Do not commit local trust state.
• Curriculum notebooks are generated by scripts/build_curriculum_notebooks.py. Update the generator, regenerate notebooks/, and keep notebook code cells output-free. Execute every changed code cell before committing. Keep the ratchet rule: each load-bearing idea gets a napkin-scale AND a real-scale runnable pair.
• The per-fork apex boundary axiom sets in profiles/ed25519.py MIRROR each verified repo's check.sh Phase 3b; the button is the authority. When a repo's enforced boundary changes, update the table in the same change set and say so.
• Apex-tier certificates are judged against their fork's documented boundary, standard certificates against the three standard axioms; deviation in EITHER direction is dirty. Never widen an allowed set to make a cone pass.
• Trust providers for OBSERVATIONS, never VERDICTS: axiom_status of attested certificates is re-derived locally from observed_axioms. Do not bypass _normalize_certificate.
• Partial attestations degrade (uncovered certificates stay unproven; the score caps); they are not rejected. Identity, signature, and receipt failures still reject.
• STH pinning: same size means same root (EQUIVOCATION otherwise), growth requires a verified consistency proof from the pinned size, shrinkage is rollback. Never accept growth without proof, and always check the consistency anchor's ROOT against the pin, not just its size.
• Dogfood verifier: never hardcode a machine path in committed files (Cargo.toml stays a template); always write the provenance sidecar; record the backend that actually verified each signature; OpenSSL fallback must be visible in evidence and must fail --require-verified-verifier policies.
• ML-DSA stays fail-closed until a real, configured backend exists. The dogfood loop does NOT extend to it: there is no proven ML-DSA, and the hybrid posture depends on saying so.
• scripts/mini_pytest.py must keep running the whole suite on a bare python3; if a test needs a new pytest fixture, extend the shim in the same change.