docs/testing/2026-09-30-eval-baseline/clankie-suite-summary.md

clankie · claude 2.1.285 (Claude Code) · claude-sonnet-5-5

ArmPass rate95% CITokens/trial (95% CI)Wall/trial (95% CI)TrialsErrors
bare75/80 (94%)86% to 97%24k (22k to 26k)8 s (7 s to 9 s)800
current80/80 (100%)95% to 100%51k (47k to 55k)8 s (7 s to 9 s)800

current − bare, paired by case (case-clustered bootstrap):

MeasureDifference95% CIVerdict
Pass rate+6%+0% to +19%within noise (16 cases)
Pass rate: heldout+0%+0% to +0%within noise (5 cases)
Pass rate: incident+0%+0% to +0%within noise (5 cases)
Pass rate: social+17%+0% to +50%within noise (6 cases)
Tokens/trial+27k+22k to +33koutside noise (16 cases)
Wall/trial−0 s−1 s to +1 swithin noise (16 cases)