Claude wrote scripts/eval_gate.py for your CI and it "worked" for
two weeks — every run green. Then you read it: the verdict on line 5
only checks pytest. A 0.71 model sailed past a 0.80 threshold
fourteen times. Fix line 5 so the run is green only when tests pass
AND eval_score clears THRESHOLD.
Expected output:
red X
The break is on line 5 — but read the whole snippet first.