Thanks for your interest. RuntimeGuard-AI is a research prototype with an evidence-first culture: every public claim must map to code, tests, or a canonical measurement. Contributions are held to the same bar.
Rust 1.92.0 is pinned by rust-toolchain.toml. The single verification
entrypoint is:
./reproduce.sh --verifywhich runs cargo fmt --check, strict Clippy, all workspace tests,
cargo audit, and the Python tooling tests. CI runs the same gates plus the
clone-to-first-policy smoke test (.github/workflows/verify.yml).
- Policies. Share policy sources (
.rgpfiles) underexamples/policies/with a comment header describing the threat they address. Seedocs/writing-policies.md. - Engine and attestor code. Bug fixes and features in
src/inline,src/attestor, andsrc/cli. Behavior changes need tests that fail without the change. - Documentation. Corrections and clarifications to
docs/and the README, holding to the project's claim discipline (below).
- Do not add claims about zero-knowledge execution, model-inference
correctness, transparency witnessing, anti-rollback under a compromised
host, or regulatory conformity.
docs/protocol-v2.mddefines the boundary. - Performance numbers in documentation must come from canonical runs under
results/with an immutable run directory; quick runs are diagnostics. - New protocol-visible fields require a schema-version discussion in the PR.
-
./reproduce.sh --verifypasses locally. - New behavior has a failing-first test.
- Public docs updated when observable behavior changed.
- No unexplained changes to
results/orpaper/artifacts.
If you find a vulnerability in the protocol implementation (signature verification, recovery validation, lease handling, frame parsing), please do not open a public issue. Email the maintainer via the contact on the GitHub profile with reproduction details.