Benchmarks

Run August 2026 · Crinkle 0.1.0-alpha.17 · raw scorecards downloadable below

The claim being tested: the same model produces more software that meets its requirements, and stops falsely claiming success, when it works inside Crinkle. One mid-size open model, three arms, graded by external browser automation that clicks the built apps. No model judges any model.

Method

Hard suite: 5 tasks × 3 runs

ArmRuns passedChecks passedFalse completionsWall timeTokens
Single-shot1/15 (7%)68/123 (55%)10/15 (67%)19 min~0.1M
Naive loop2/15 (13%)79/123 (64%)9/15 (60%)21 min~0.1M
Crinkle8/15 (53%)107/123 (87%)06.9 h3.95M

Where the bare model failed: it dropped features from multi-feature specs, broke working behavior during the refactor, leaked NaN under interacting hostile input, implemented the scheduling state machine loosely, and fixed one of three planted bugs while declaring victory. Crinkle's two 0/3 tasks ended at 23/27 checks each, and it never claimed success on them.

Standard suite: 17 tasks × 3 runs

ArmRuns passedChecks passedFalse completionsWall timeTokens
Single-shot42/51 (82%)152/168 (90%)8/5120 min0.12M
Naive loop46/51 (90%)161/168 (96%)5/5122 min0.14M
Crinkle51/51 (100%)168/168 (100%)09.5 h5.67M

Honest caveats

Raw scorecards

Every number above traces to these files, exactly as the harness wrote them (per-run checks, end states, criteria verification detail, per-agent call counts):