crucible-bench is distributed on PyPI. Security fixes go into the current release line as a new release. Older lines are not backported.

1.2.0 and earlier carry the verdict-integrity defects fixed in 1.3.0: a measurement could widen a claim's tolerance after the seal and still re-derive MATCH, and doctor reported MATCH for checks it never ran. See the 1.3.0 entry in CHANGELOG.md. Upgrade with pip install --upgrade crucible-bench.

Report suspected vulnerabilities privately through GitHub: open the repository's Security tab and choose Report a vulnerability, or go straight to https://github.com/HarperZ9/crucible/security/advisories/new. Do not open a public issue, pull request, or discussion for a security report.

Include:

• the affected version and platform,
• what the issue lets someone do,
• the smallest input or steps that reproduce it.

You will get an acknowledgement, and the fix ships before any public disclosure. Please allow a reasonable period to remediate before you disclose.

crucible-bench is a judgment engine. It registers a thesis, measures it, and emits a verdict per claim that anyone holding the sealed record can re-derive. The core has no runtime dependencies and makes no network calls.

• What counts as a vulnerability here. Anything that lets a verdict be accepted that the sealed record does not support: a MATCH that survives re-derivation without a measurement earning it, a tampered record that still verifies, or a check that reports success without running. Path escapes from a registry and code execution from data files count too.
• Child processes. The crucible command and the MCP server start no child process. The library edges ProofMeasure and SubprocessMeasure do: they run a command you configure, with your privileges, in the working folder of the calling process. Only configure commands you trust, and run measurements that execute external code in a container or other sandbox.
• Untrusted input. Thesis files, measurement records, and substrate numbers may come from outside sources. crucible treats them as data. The sealing layer is built to fail on tampering; it does not make a supplied measurement true.
• Writes. Commands write only where their arguments point, and refuse to overwrite an existing output file or bundle folder.

• Keep API keys and backend credentials in the environment, never in a thesis, measurement file or committed record.
• Seal the tolerance on every new claim ("tolerance" in the thesis file). A claim without a sealed tolerance keeps the legacy behavior, where a widened tolerance can still re-derive MATCH.