For coding-agent task authors

Prove a coding task works before an AI sees it.

RepoGauntlet confirms the broken code fails, an incomplete fix is rejected, the reference fix passes, and repeated runs produce the same normalized outcome.

5languages

5task packs

15controls

PY-INTERVAL-001BUG FIXMedium

Repair half-open interval reconciliation

Merge unordered half-open windows, reject reversed intervals, discard empty intervals, and preserve the caller’s input.

Calibration proof

Select a control to inspect its saved evidence

Broken baselinemust failDefect detectedIncomplete fixmust failShortcut rejectedReference fixmust pass ×3Task is solvable

Saved evaluation run

RESOLVED

seed: 42 / golden

✓ build passed 92ms

✓ public tests passed 215ms

✓ held-out tests passed 218ms

✓ quality passed 86ms

score 100.0 / 100 · outcome fingerprint ec4da2de4524…

Phase results

build10 pts

public tests25 pts

held-out tests65 pts

quality0 pts

611ms

Recorded time

58c59b1

Bundle fingerprint

Inspect source, tests, report, and CI

Task manifest Grader tests Selected code Raw report Polyglot CI

Candidate overlay accepts only solution filesRejects path escapes, symlinks, and oversized overlay files before execution

Repeatability checkThree reference runs must share one outcome fingerprint

Separate quality gateDistinct checks earn points; redundant checks earn zero