evals

One exam, one seat. A reviewer miss is the expensive kind, and a diff with a known defect is a cheap, honest exam of exactly that. Every other seat (brainstorm, plan, execution lanes) is judged on trial runs: the worklog's Run:/Seats: lines, compared by checkup.

Run it before giving a new model the strict-reviewer seat, and before keeping a change to workflow/references/phase-4-review.md (same model, old and new instructions). The human invokes it.

The exam set

One central set for every repo, in the private repo j-skills-evals, cloned at the fixed path ~/.agents/j-skills-evals, so the cases are in ~/.agents/j-skills-evals/strict-reviewer/. The path is fixed because installed skills are copies, not git checkouts, so no lookup from the skill folder works. Missing → stop and give the human the clone command: gh repo clone johnviklund/j-skills-evals ~/.agents/j-skills-evals. The set is private because cases copy code from private repos; never copy a case into a public repo.