Measure posture, differences, and evals

Capture a baseline posture

Run the posture workflow against the secure lab control with the same local-target safeguards used in the scan lesson. Record version, configuration, and timestamp.

basilisk posture --help

Compare controlled runs

Use basilisk diff to compare saved runs that share provider, model, module set, and request policy. Do not attribute unrelated configuration drift to the target.

basilisk diff --help

Create an eval expectation

Build a small eval containing one vulnerable expectation and one secure expectation. Keep prompts deterministic and make pass/fail assertions evidence-based.

basilisk eval --help

Interpret regressions

Treat a changed result as a signal to inspect. Authentication failure, endpoint drift, nondeterministic output, and tool errors are not automatically security regressions.