Glossary · Machine learning
Evaluation gate
Also called gate, the gate, evaluation, deterministic evaluation
A fixed set of tests a trained model must pass before it is kept: here every skill, played without exploration noise, against fixed seeds.
During training, a policy’s scores are noisy. It is sampling random actions to explore (see stochastic policy), the situations are random, and the averages lag behind. An evaluation gate is a separate, fixed test: the same scenarios and seeds every time, the policy played without noise, and a pass mark for each row. A checkpoint is kept only if it passes.
A gate measures only what it tests. A model can pass it and still be wrong in ways nobody wrote a test for, which is a cousin of reward hacking.
In this project
The skills gate plays every skill in fixed scenarios. Punches are 120 presses in the air and at a bag. Walking is 40 goals, blocks 32 presses (with and without shoves), the opponent 160 punches, and distances 60. Each new technique has its own eval, with hits within 5 cm at hitting speed, falls, and since 2026-10-26 the hips’ turn after the technique. The gate swings from snapshot to snapshot more than the training averages do (blocks between 21 and 29 over distill3’s snapshots), so several snapshots are gated and the best is kept. Several teachers passed it and were still judged failures on video: rolls that drop instead of rolling, and a lead uppercut that ends turned 90° (see Shadow clones).