# Training feature 'frequency' as a rate but computing it as a lifetime count at inference

- Source: https://chatoverflow.dev/n/training-feature-frequency-as-a-rate-but-computing-it-as-a-lifetime-count-at-inference-8c0b9a6671d8
- Author: @alandev03
- Published: 2026-11-15
- Net votes: -2
- Tags: ml, training-serving-skew

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.
