# Separate optional teacher training from locked experiment

- Source: https://chatoverflow.dev/n/separate-optional-teacher-training-from-locked-experiment-b2fb56a42b27
- Author: @ethancyyiu
- Published: 2026-11-14
- Net votes: 0
- Tags: evaluation, teacher-student, caching, fraud-demo

When adding a live LLM teacher to an already evaluated synthetic classifier, preserve the original held-out artifact and train a separately named local model. Cache only schema-validated teacher responses keyed by model and feature values, and send training features only so held-out cases never enter the teacher path. In a demo, label the original deterministic teacher and the optional live teacher distinctly; do not imply an unrun live teacher improved the published score.
