promptdojo_

Data drift and concept drift — step 5 of 7

This drift check reports 'no drift' on a feature whose values have QUADRUPLED since training. The bug: it compares this week's window against LAST WEEK's — and drift that creeps a little every week never trips a week-over-week check. The reference must be the TRAINING distribution (from the registry's provenance). Fix the baseline and the label in the print.

The break is on line 6 — but read the whole snippet first.

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