This project makes three kinds of guarantee, and a failure of any of them is a security defect rather than a bug:

1. Suppression guarantees. A source claiming implausibly broad activation must not produce alarms. If you can construct a feed state that raises an alarm through the suppression, that is a defect.
2. State guarantees. No state other than CLEAR may read as an all-clear, and none other than ACTIVE may drive an alarm. A path where UNKNOWN behaves as CLEAR is a defect, and so is one where PARTIAL_CLEAR does: a message announcing an all-clear while saying the alert continues is a contradiction, not an all-clear. The guarantee is written over the state enumeration rather than over a list of state names, so it covers a state added after this sentence.
3. Claim guarantees. Every limitation in the README is registered in tests/lint_limitations.py. A claim in the documentation that the tree does not implement is a defect, and historically the most likely one.

Also in scope: any input that causes a parser to raise, and any path that writes a raw per-subject record outside data/raw/.

Open a GitHub security advisory on this repository. If the finding concerns a guarantee above, please include the input that demonstrates it; a probe is worth more than a reading.

Expect an acknowledgement within seven days. This is a single-maintainer repository, so a fix is not promised on a schedule, but the defect will be recorded in docs/METHODOLOGY.md with its class either way.

The correctness of upstream feeds. This tool reports what sources say, labelled as reported, and cannot verify their claims.