# Browser-agent evaluation checklist

## Contract
- [ ] Outcome is observable and bounded.
- [ ] Every expected source is named.
- [ ] Allowed and prohibited effects are explicit.
- [ ] A missing source becomes `UNKNOWN_RECONCILING`.

## Execution
- [ ] API or deterministic route attempted before general browser reasoning.
- [ ] Capability lease is narrower than its parent.
- [ ] Credentials are origin-bound and absent from model context.
- [ ] Unknown writes are reconciled before retry.

## Verification
- [ ] `DONE` is not accepted as proof.
- [ ] At least one current authoritative evidence path passes.
- [ ] False-success negatives include stale pages, partial lists, wrong account, previous-run banners and timed-out writes.

## Promotion
- [ ] Evidence matches the exact workflow, site, model, browser, prompt, policy and verifier version.
- [ ] Repeated-run consistency is measured.
- [ ] Drift and OOD cases are represented.
- [ ] Promotion decision is `PROMOTE`, `HOLD` or `DEMOTE`.