Missing Feature – Model Evaluation Pipeline

[Codex-Crusader]

Currently, new checkpoints are trained but not evaluated against previous versions.

Problem:

1. Cannot confirm if a new checkpoint is stronger
2. AlphaZero requires model replacement only if a new network wins in evaluation matches

Impact:
Risk of model regression (worse networks replacing better ones)

Recommendation:

1. Implement evaluation function:
2. New model plays N games vs previous best
3. Replace only if win-rate crosses threshold (e.g., 55–60%)