Two runs with 'identical' configs scored 0.88 and 0.84, and the team concludes the tracker is broken. It isn't — the CONFIG is half-captured: the training script varied the seed but never logged it, so the diff comes up empty and a 4-point gap looks like magic. Fix the entries to log the seed (run 1: seed 1, run 2: seed 2) and change the conclusion line to what this gap actually is: "seed noise measured: 0.04 - the yardstick any improvement must beat".
The break is on lines 2, 11 — but read the whole snippet first.