pagekit-evaluator-pass

PageKit Evaluator Pass

You are invoking the adversarial evaluator pass on a completed run.

When to use

• After the run's evaluation.md is written.
• Before declaring the run fully-logged (the fully-logged tier requires evaluator-pass.md).
• When a reviewer asks "is this run as good as the evaluation says?"

Read first

• ../pagekit/references/run-logging.md — where the evaluator pass fits in the run-logging tier
• agents/pagekit-evaluator-pass.md — the subagent you will delegate to

Inputs

• path to the run folder (e.g., runs/kind-bowl-real/)

Procedure

1. Confirm the run has evaluation.md. If not, the evaluator pass is premature — finish the evaluation first.
2. Invoke the pagekit-evaluator-pass subagent. Pass the run-folder path.
3. The subagent runs read-only, produces runs/<name>/evaluator-pass.md, and hands back a summary (count of scrutiny items, count of punch-list items, one-line final read).
4. Relay the result to the caller. If the punch list contains actionable repo changes, name them for the caller so the run-to-repo-improvement loop can close.

Quality gate

A strong evaluator pass:

• names specific quiet claims from the run, not general impressions
• writes an "implication" for each scrutiny item (what this suggests for the method or the repo)
• ends with a concrete punch list of file-level changes
• does not reward defensiveness

If the subagent's output is general rather than specific, ask it to retry with more direct reference to the run's artifacts.

Relationship to pagekit-claim-checker

• pagekit-claim-checker reviews the draft line by line at a chosen severity.
• pagekit-evaluator-pass reviews the whole run including what the evaluation quietly claims. Different target, different voice.