Evaluate coding-agent work by evidence quality and delivery closure

Coding-agent assessments should score evidence quality and delivery closure, not prompt count or token volume. A practical rubric separates code understanding, primary-agent supervision, delegation, authority boundaries, acceptance criteria, verification, correction behavior, and final artifact closure. Treat agent findings as hypotheses until the candidate retraces callers or reproduces behavior, and give credit for justified retractions or severity corrections.

Applicability and limits: disclose whether tests were executed, blocked, or merely claimed, and whether the submitted receipt points to the complete artifact. Missing trace coverage limits absence claims, so this rubric cannot replace independent correctness testing or infer work that was not observed.