Review is against an artifact, not a summary
An agent that describes its own change is not evidence. The reviewer reads the artifact that would ship, at the revision proposed.
A filled review
Ridgeline Roofing, one proposal, reviewed 2026-09-11.
proposal rewrite the title and description on three service pages
artifact three files shown in full, one revision pinned
evidence search console, 28 days, two queries per page
tier measured
prediction clicks, up, over 28 days, direction only
preview built from the same commit, audited, page renders
approver Sam, owner, recorded 2026-09-11 16:40
reviewer checks
1. the change names the three files yes
2. the artifact is shown in full, not summarised yes, three files
3. evidence carries a source and a tier yes, measured
4. the prediction names metric, direction, window yes
5. the preview was audited yes, three pages read
6. the decision records who approved it yes, Sam
7. nothing shipped while the review was open yes
decision approve, with one revision: keep the phrase "roof repair"
in the first title, because that is what customers say
on the phone
What the path refuses
- a proposal that is not approved
- a proposal approved with no recorded decider
- a proposal that has already shipped
- a measurement window that disagrees with the prediction
- a failed preview audit
- no measurement source connected to read the result
Those refusals are in the code, not in a prompt.
Acceptance: the review is complete when
- the artifact is read in full at the pinned revision
- evidence carries a source and a tier
- the prediction is falsifiable
- the approval records a person
- the decision and any revision are written down together
Troubleshooting
| Symptom | Likely cause | What to do |
|---|---|---|
| The agent summarises its own work | No artifact was requested | Ask for the files, not the description |
| The change is approved with no name | A tool recorded it | Refuse it, and re-approve properly |
| The preview was not audited | No preview URL was resolved | Link the repository, then re-run |
| The prediction has no metric | It was written as a wish | Ask for metric, direction and window |
| Nothing can be measured | No source connected | Fix access before shipping |
Limits
Review catches what a person can see. It does not guarantee the change was a good idea, which is what the measurement window is for.
An inconclusive result is still a valid outcome, and it is worth recording.