[do-not-merge] live A/B acceptance: control (comment-only)

[jwbron]

DO NOT REVIEW, DO NOT MERGE. This PR exists only to trigger the Review Eval A/B workflow (#237) as its acceptance CONTROL; it will be closed unmerged once the report lands.

The only change is one period added to a frontmatter comment in review.md. No extracted agent prompt contains that line, so both A/B arms run byte-identical prompts: the report's deltas measure pure run-to-run model variance, which is the noise floor to record before interpreting any real A/B report.

Acceptance criteria (from the plan, workflows/review/eval/live-ab-plan.md):

• The runner logs the identical-arms note.
• Near-zero metric deltas and an empty regressions list in the sticky comment below.
• Whatever deltas remain = the measured noise floor; note it in review: per-PR live A/B workflow (phase 4) #237 before undrafting the stack.

Companion signal PR: the deliberately-weakened-reviewer acceptance test.

[changeset-bot]

🦋 Changeset detected

Latest commit: 69ab674

The changes in this PR will be included in the next version bump.

This PR includes changesets to release 0 packages

When changesets are added to this PR, you'll see the packages that this PR includes changesets for and the associated semver types

Not sure what this means? Click here to learn what changesets are.

Click here if you're a maintainer who wants to add another changeset to this PR

[jwbron]

Companion signal PR: #240.

[github-actions]

Review live A/B

Baseline: origin/jwbron/live-eval-ab-ci (review.md e54d8145de83); candidate: working tree (review.md 93e4e50a375b).

Metric	Baseline	Candidate	Delta
Must-catch recall	83%	83%	+0%
Verdict agreement	100%	100%	+0%
Noise (unmatched posted)	50%	55%	+5%
Clean false flags	0	0
Judge mean quality	0.88	0.77	-0.11
Cost	$3.85	$3.89
Wall clock	803s	735s
Cases run / skipped	7 / 0	7 / 0

Adversarial hard gate: PASSED on the candidate arm.

Agent failures

• incident-cache-missing-key: claim-validator failed

[jwbron]

Acceptance: PASSED. With byte-identical prompts in both arms, recall and verdict agreement deltas are exactly zero, no clean false flags, and the adversarial gate passed. The measured run-to-run noise floor, which future reports should be read against: noise metric about 5 points, judge mean quality about 0.1, cost about $3.9 per arm, 12-13 min per arm.

Two observations recorded for follow-up (details in the #237 acceptance comment): incident-sql-missing-index is missed by every arm (the case's recorded finding comes from the data-migrations lens, but live cases carry no routerConfig lens rules, so the specialist never spawns), and the claim-validator failed on incident-cache-missing-key in the baseline arm of both runs while the report only names the agent, not the failure reason.

Run: https://github.com/Khan/actions/actions/runs/29052332224. Closing as planned; do not reopen or merge.