How-tos

Define a small AI evaluation

Create fixed ordinary, absent-answer, and hostile-content cases with observable pass criteria and named revisions.