A structured way to review a system for clarity, usefulness, safety, and learnability. Companion to Lesson 90.10. Hand it to any reviewer — human or AI — with the instruction: challenge gaps, do not flatter.

The rubric (score 0–2 per row: 0 missing, 1 asserted, 2 evidenced)

# REVIEW-RUBRIC — [Project], reviewer: …, date: …

## Clarity (can a stranger picture it?)
- [ ] Person + job named (0/1/2) — evidence: …
- [ ] Outcome + proof observable (0/1/2) — evidence: …
- [ ] Non-goals explicit with triggers (0/1/2)

## Usefulness (does one path work?)
- [ ] One flow completes for one user (0/1/2) — preview result: …
- [ ] Acceptance criteria testable + checked (0/1/2)
- [ ] Route + action + exchange designed in (0/1/2)

## Safety (what stops harm?)
- [ ] Agent boundaries: spectrum + gates + never-list (0/1/2)
- [ ] Human owns consequential judgment + publish (0/1/2)
- [ ] Logs + kill switch + rollback named with owners (0/1/2)

## Learnability (will reality teach?)
- [ ] 3 hypotheses with 4-signal mix + dates (0/1/2)
- [ ] Cost caps + quality sample + support route (0/1/2)
- [ ] Single decision asked, with decider + date (0/1/2)

## Totals: Clarity __/6 · Usefulness __/6 · Safety __/6 · Learnability __/6
## Verdict: [ship pilot / fix-list (tasks: …) / re-scope]
## Gaps → tasks: [gap → task → owner → date]

How to run the review

Score with evidence pointers, not impressions: each 2 must cite a file and line ("AGENT-BOUNDARIES §gate, owner Maya"). Any 0 blocks a pilot; any three 1s become the fix-list. Ask the six questions from Lesson 90.10 alongside scoring: real problem? untested assumption? unnecessary component? harm if wrong? approval needed? how known? Record the hardest question and the change it caused in the review log.

Passing bar for a controlled first release: no zeros, Safety 5+/6, and a named preview user plus cost cap. Anything below that is a fix-list with owners and dates — not a rejection, a work plan.

Calibrating scorers

Two reviewers scoring the same plan should land within two points per section or the evidence pointers are vague. When scores diverge, the lower scorer names the missing artifact line and the higher scorer either produces it or concedes the point. Run this calibration once with an AI reviewer and once with a domain human: the AI finds unlogged permissions and untestable criteria, the human finds false jobs and unwanted promises. Keep both score sheets in the review log.

Verify fast

The outsider test: a stranger reading only FINAL-PROJECT-REVIEW.md states the user, outcome, proof, cost cap, exclusions, and the one ask. If any answer fails, the corresponding rubric row was overscored. Correct the score, add the task, re-review the row — not the whole plan.