feat(eval): score the reviewer on past culprits and clean PRs

AGENTS.md

@@ -24,6 +24,10 @@ Do not edit `CONTRIBUTING.md`, `assets/CONTEXT.md`, `assets/AGENTS.md`, or
   PR, or the standing one /quest-run gives a PR with no user-facing behavior
   change that adds or expands no quests; and only after repository checks and
   reviews pass.
+- A change to the reviewers in `src/cli/models.ts`, the review prompts in
+  `src/cli/review.ts`, or a `REVIEW.md` runs `quest eval run` with the base's
+  build and the head's, and puts both scores and their cost in its PR's
+  `## Validation`.
 - Pull requests land as squash merges.
 - Do not comment on repositories outside kixelated/moq-dev without approval.
 

README.md

@@ -1558,6 +1558,33 @@ the reason of the first to judge it so. The record names it after the base:
 tier, and a reviewer that raised it adds `- Raised the risk to high: <reason>`
 to its block.
 
+### Score the reviewer
+
+`quest eval` scores the reviewer on the repository's past PRs, kept in
+`review-eval.json` at its root: culprits, each with the places at its last head
+of the defect a later fix's `## Caused by` named, and clean PRs that broke
+nothing. Score a change to the reviewer table, the prompts, or `REVIEW.md`
+before it lands, and put the score and its cost in its PR.
+
+```sh
+quest eval add 570 --fix 704 --what "It counts every link as a culprit." \
+  --defect src/cli/usage.ts:395-399 --defect src/cli/usage.ts:408-410
+quest eval add 503 --clean
+quest eval run                          # the size's pick and its shadow on every case
+quest eval run --harness claude --effort high --pr 570
+```
+
+`add` reads the PR's last head and base from GitHub and records its merge base,
+size, latest writer, and tier, and a culprit already in the set gains each
+further defect. `run` fetches each case from the set's repository, so a fork's
+clone scores its upstream's set, and reviews it at its head in a detached checkout,
+giving the reviewers the first round `quest review` gives, with the base's
+`AGENTS.md` and `CLAUDE.md` and this checkout's committed `REVIEW.md`. A
+finding on any line of a defect's places finds it; every finding on a clean PR
+is noise. It prints defects found and clean findings by severity for the pick,
+its shadow, and either, by size and by tier, each case's rounds, and the cost
+token-max prices for the run's sessions.
+
 ## Audit cost and accuracy
 
 `quest usage` reports `origin`'s PRs updated since `--since` (a local

quest/m0/mistake-audit.md

@@ -66,6 +66,11 @@ Agent's choices, for the implementer to revisit. Each run:
   owns that gate, so it goes in the run PR's `## Follow-ups` for
   `/quest-plan` there, as the guide's Creation section says (user,
   2026-11-11, [routing evidence](/quest/m0/routing-evidence/README.md)).
+- Add each escape classified catchable or maybe to its repository's eval
+  set with `quest eval add`, its defect located at the culprit's head from
+  the fix's diff and confirmed against the culprit's own diff, so `quest
+  eval run` scores reviewers on it. A private repository's set stays in
+  that repository.
 - Before proposing a fix, find the quest, check, or rule that already owns
   the class, and report whether its count fell since that landed. A rule
   whose class still recurs is the case for moving it up a level.

quest/m1/model-check.md

@@ -42,8 +42,9 @@ checkpoints and outages to count by hand:
 - Move a size to a stronger model or effort where its PRs needed clearly
   more rounds or fixes than the size above, and to a cheaper one where its
   reviews come back clean. Do the same for the review efforts in `src/cli/models.ts`
-  from what later reviews and follow-ups found that a review missed. Record the tallies in
-  the next run's Plan as its baseline.
+  from what later reviews and follow-ups found that a review missed, and score each
+  reviewer change with `quest eval run` before recommending it (`README.md`'s Score the
+  reviewer). Record the tallies in the next run's Plan as its baseline.
 - Recommend each change to a list's order, harness, or effort to the
   maintainer with its tally, as a decision in the PR, rather than changing
   it unasked. Replacing a stale model id with the same vendor's newer one,