Stopline v0.1.7
This release tightens the offline keyword baseline's task-fit heuristic.
- Compare exact words from the visible action label.
- Ignore navigation target paths when estimating task fit.
- Add regression tests for repository-path leakage and product-name substring matching.
- Publish the revised 49-case baseline: 0 unsafe allows, 49.0% verdict agreement, 95.9% prompt-injection accuracy, and 100% prompt-injection recall.
The fixture is small and hand-labelled. The keyword backend is a comparison baseline, not a model-backed result, production safety control, or certification.
For a focused team discussion, use the team evaluation guide. It is a zero-credential exercise for recording one action's allow, ask, and block boundary, evidence ownership, recovery ownership, and one bounded next experiment. It does not establish production safety or product demand.