Security fixes are applied to the current development line. Frozen research tags and exact experiment revisions remain immutable for reproducibility; a security issue affecting one of those revisions is documented and repaired in a new version rather than silently rewriting the historical source.

Do not disclose a suspected vulnerability in a public issue, pull request, log, or artifact.

Use GitHub's private vulnerability-reporting or security-advisory interface for this repository. Include:

• the affected revision or package version;
• the smallest reproducible example;
• the expected and observed behavior;
• whether credentials, private data, self-hosted runners, artifact integrity, or arbitrary code execution are involved; and
• any temporary mitigation already applied.

Do not attach restricted datasets, credentials, access tokens, private model weights, or third-party material unless its terms permit that disclosure.

The following changes receive heightened review:

• GitHub Actions and self-hosted runner execution;
• artifact loading, path validation, archive extraction, and deserialization;
• content-addressed identities, evidence locks, and provenance validation;
• cross-repository provider admission;
• credentials, data roots, model caches, and external downloads; and
• code that changes fail-closed behavior or exact fallback.

Pull-request code must not receive private dataset roots, confirmation payloads, SSH keys, package-publishing credentials, or a writable Docker socket merely to run tests. Privileged empirical workflows should execute only reviewed exact revisions and should keep dataset-bearing execution separate from general pull-request validation.

A report is first triaged for reproducibility and impact. Confirmed issues are assigned a private repair plan and a disclosure boundary. Research provenance is preserved: affected evidence is marked, successor artifacts receive new identities, and historical result bytes are not rewritten.