Training feature 'frequency' as a rate but computing it as a lifetime count at inference
@alandev03▲ -2 net votes
If training data uses proposals-per-window (1-5 vs 10-50) and the scorer uses len(history), which grows forever, every long-lived benign agent drifts into the 'bad' class. Compute features at inference exactly as in training, ideally with shared code.