self-reviewlisted

# Self review Reviewing your own work in the context that produced it does not work. You still hold the reasoning behind every line, so you read the intention rather than the code — and the reviewer who receives it will have only the code. Two things fix that. Audit the evidence rather than trusting the memory of having produced it, and hand the diff to reviewers who never saw it being written. ## 1 — Audit the evidence Before reading any code. For each acceptance criterion in the spec, find the artefact that proves it and ask whether it actually proves it: | Criterion says | Evidence is | Verdict | |---|---|---| | Error panel appears on 403 | `orders-error-403.png` showing the panel | Proven | | Error panel appears on 403 | The unit test suite passed | **Not proven** — nothing exercised the 403 path | | List renders within 2s at 500 rows | No artefact | **Not proven** — record as unverified | An unattended run accumulates evidence that looks complete and covers the happy path. This audit is where that becomes visible, and it is cheap: a criterion with no artefact or with an artefact that proves something adjacent is **not done**, whatever the code looks like. Confirm the blast radius while you are here: run the changed-set command from `docs/agent/repo-map.md` and compare against the spec's expected set. ## 2 — Dispatch three reviewers Three subagents in parallel, each in **fresh context** — they see the diff and their own brief, never the reasoning that produced th