Prove · Evals

Evals

An eval bar is a verifier promoted from a per-loop command into a named, versioned bar: a case dataset plus a pass threshold. The bar is versioned, so raising it is a recorded event, not a silent edit. The Evals tab shows your verified memory as a leaderboard: which bars exist, the pass rate over each run, which loops clear which bar, and where a regression crept in. RepoOps runs no loop of its own. It scores the proof output your loops already produce, and a case with no proof reads an honest "not evaluated", never a false pass.

Where to find it

  • Local tab: /evals.html?repo=<id>
  • List bars API: GET /api/eval-bars?repo=<id>
  • One bar API: GET /api/eval-bars/<id>?repo=<id>
  • Bar artifact source: .claude/eval-bars/<id>.md
  • Run log: .claude/brain/eval-bars/<repoId>.jsonl

What it does for you

A leaderboard of your bars, ranked by pass rateEvery eval bar for the selected repo lists in one column, sorted with the highest latest pass rate first. Each row shows the latest pass rate, the bar's threshold and version, a badge for whether the loop clears or falls below the bar, case count, and run count. Pick a bar to open its detail.
Pass rate over time, against the barThe detail view draws a sparkline of the pass rate across the run series with a dashed line at the threshold. Green dots are runs that cleared the bar, red dots fell below. Each dot carries the run's timestamp, pass rate, and version, so you can read when a loop crossed the bar.
Which loops clear which bar, and where a regression crept inA table lists each loop that ran against the bar and whether it clears or sits below. A separate panel flags regressions: a loop that once cleared the bar and later fell below it, showing the drop from one pass rate to another at the version where it slipped.
Versioned thresholds you can trustRaising the bar is recorded. The threshold-version table lists every version, its threshold value, and when it was set, and older runs stay attributed to the version that was live when they ran. The latest run's per-case verdicts (pass, fail, or not-evaluated) round out the honest picture.

Read more