promptdojo_

Train/inference skew — step 6 of 7

Write skew_check(probes, tolerance=0.01): push every probe value through BOTH normalizers, find the worst absolute gap, print f"worst gap over {n} probes: {worst:.3f}", then "PASS" or "FAIL: serving path diverges - do not deploy". Run it on the five probes. This is the smoke test that belongs in front of every deploy — testing through the SERVING path, not the training one.

full-screen editor opens — close anytime to keep reading.