experiment

You can run an iterative optimization loop that proposes changes, measures a single target metric, and uses a ratchet pattern to keep improvements while reverting regressions. This requires a methodology file defining your numeric target, direction, and measurement protocol.

A ratchet-based optimization loop commits changes, measures outcomes against a target metric, and automatically reverts regressions. It enforces strict tie handling and git commit discipline, ensuring only validated improvements are kept while maintaining a clean lab notebook trail.

You log experiment iterations to a structured `results.jsonl` file that records outcomes including skips, correctness gate results, and reasons. This JSONL logging conforms to the iteration protocol and supports resuming experiments from existing results.

No, you must use a dedicated worktree path rather than your main branch. The workflow enforces git-safe experimental discipline by committing changes and reverting discards, ensuring a clean lab notebook trail and protecting your main branch from experimental noise.

Use futility stopping to halt the optimization loop after consecutive unproductive iterations, preventing wasted computation when the metric stops improving. This prevents wasted computation when the measurable metric plateaus, while still allowing the experiment to resume from existing JSONL results later.