ralph-experiment

Structured hypothesis-driven experiments require a baseline measurement, a Git-backed notebook schema, and an agent-enabled delegation workflow to isolate iterations and log learnings. You define a hypothesis, run a baseline, and iterate one change at a time.

Git checkpoints support reverting failed experiments during performance tuning by keeping each iteration isolated. You can safely test a prediction, capture the result in a notebook, and revert the codebase if the experiment does not improve the baseline.

Agent delegation distributes experiment work across executors, analysts, and architects. This workflow automates the iteration loop so agents handle tasks like running measurements and analyzing results, which are then tracked in the experiment notebook.

A baseline measurement is required before starting feature exploration iterations. It provides the reference point needed to evaluate whether each hypothesis-driven change actually improves performance or reliability.

The best way to track learnings across experiments is using a cross-iteration notebook combined with Git-based state tracking. This captures insights and guides next steps by logging hypotheses, predictions, and results throughout the iteration loop.

You should not use a hypothesis-driven experiment loop for workflow automation if you lack a baseline measurement or cannot isolate changes with Git checkpoints, as structured iteration requires tracking state and reverting failed tests.