Shipped in three parts on the review-bot side:

zuiho-kai/omni-reviewbot#15 — every review job now records what it produced: severity histogram (P0–P3), inline anchor rate, gate failures, posted review ids. The dashboard shows the aggregates; NULL renders as "未知", never as 0, and every rate is displayed with its sample size.
zuiho-kai/omni-reviewbot#19 — the "was it good" half: a daily read-only collector gathers 👍/👎 reactions on the bot's inline comments, thread replies, and each PR's final state. Maintainer convention: react 👍/👎 on the bot's inline comments — that is the signal the dashboard's useful-rate is computed from (GitHub doesn't support reactions on top-level review bodies).
eval: monthly review-quality probe of omni-reviewbot as a harness arm #109 — a monthly ground-truth probe: bash eval/dataset/run_reviewbot_monthly.sh <YYYY-MM> re-reviews the 15 train+val benchmark PRs and judges them blind against the pinned CC+Opus 5 baseline (paired, item-clustered statistics; results in eval/dataset/results/reviewbot/). This answers "is the review quality holding" with actual recall/precision deltas, monthly.

Closing as implemented. Live numbers appear on the kanban as reviews flow; the first monthly probe report will land under results/reviewbot/.

[Feature] provide review result metrics #96

Description