checkpoint-review

A/B diff evaluation analyzes Model A/B code differences alongside execution evidence and prior turn evaluations to determine a winning implementation. It produces a comprehensive, multi-section evaluation file declaring which model better satisfies the acceptance criteria.

Automated code review for A/B testing outputs compares PR-like diffs across two model implementations, applies multi-axis evaluation across execution evidence, and runs internal validation before writing a final review file. This prevents signing off on isolated changes by consuming prior turn evaluations.

To structure workspace turns for diff analysis, place your turn N diffs and execution evidence under a dedicated turn_{N} directory. The evaluation workflow then consumes this structured directory data to generate the final turn_{N}_evaluation.md file.

The evaluation workflow consumes prior turn evaluations when available to avoid sign-off on isolated changes. It collates prior evaluations alongside diffs and execution traces to produce a comprehensive final evaluation, ensuring robust model selection.

Avoid automated diff analysis when your workspace lacks execution evidence or when evaluating isolated changes without prior turn context. The workflow relies on collating diffs, execution traces, and prior evaluations to run its internal validation gate before writing the final evaluation.