Internal mechanistic evaluation not final

Analyze the model's internal workings to evaluate the process itself rather than the final output, preventing reward errors. (

Mechanistic interpretability,

AI Evaluation)