eval-implementation

To run A/B handoff experiments, you orchestrate two parallel worktrees from a common source branch, execute compare_implementations.py to generate metrics, and dispatch blind graders to produce a final structured report.

Blind grading in end-to-end evaluation dispatches automated graders with a predefined rubric to score parallel implementation sessions without knowing which strategy was applied, ensuring unbiased result aggregation and a final verdict.

Git worktree management for parallel sessions involves initializing two distinct worktrees from a selected source branch, propagating a plan file and handoff document to both, and tracking state via an eval state JSON for reproducibility.

You can automate the evaluation process by running a setup phase to initialize worktrees and propagate plans, followed by a compare phase that aggregates blind grading scores and generates a final audit-ready report automatically.

No specific dependencies are required to run parallel A/B handoff evaluations, as the workflow manages worktrees, plan propagation, and report generation natively through its internal orchestration and grading scripts.

Use A/B handoff experiments when you need to compare implementation strategies from a common source branch with controlled variables, requiring blind grading and reproducible metrics to determine the superior approach for complex handoffs.