promptdojo_

The ML-system option — six boring stages that get a repo taken seriously — step 5 of 7

This gate is a rubber stamp. It reads training accuracy — and an overfit model always aces the data it memorized. The 0.99 sails through while the held-out score, the only number that predicts real-world behavior, says 0.71.

Fix the gate so it reads the held-out score. The overfit model should be held, not shipped.

Expected output:

gated on: 0.71
gate: hold
The break is on line 8 — but read the whole snippet first.

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