promptdojo_

Train/inference skew — step 5 of 7

The skew smoke-test fails: training and serving disagree by up to 0.6 standard deviations on probe inputs, because serving hardcodes constants someone copy-pasted from a notebook cell (rounded, naturally). Fix it the structural way: one normalize(v, stats) function, fitted stats in a saved dict, and BOTH paths calling the shared implementation.

The break is on line 9 — but read the whole snippet first.

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