experiment-looplisted

# Experiment Loop Pattern ## Overview The experiment loop is a systematic approach to iterative code improvement where each change is measured against a baseline metric and kept only if it improves the result. Inspired by [karpathy/autoresearch](https://github.com/karpathy/autoresearch) and generalized for any measurable code quality metric. ## The Core Pattern ``` 0. VALIDATE → Prove the tool exists and actually RAN. No valid run, no metric. 1. DEFINE → Choose a measurable metric, its noise floor, and constraints 2. BASELINE → Measure current state; pin tool path + version + accepted exit codes 3. MODIFY → Make one targeted change 4. MEASURE → Re-run step 0 FIRST, then re-evaluate the metric 5. GUARD → Run build + tests. If they fail → REVERT, whatever the metric says 6. DECIDE → Keep only if the gain EXCEEDS the noise floor; otherwise revert 7. LOG → Record the change, the metric, and BOTH gate results 8. REPEAT → Go to step 3 (until done or plateau) ``` Steps 0, 5 and 6 are the difference between an improvement loop and a loop that optimizes for breaking the toolchain. They are specified in **Mandatory Preconditions** below and are not optional. ## When to Use This Pattern **Good fit:** - Reducing lint warnings or type errors in a codebase - Improving test coverage for a module - Reducing bundle size or build time - Optimizing database query performance (measurable via EXPLAIN ANALYZE) - Improving accessibility scores (Lighthouse, axe) - Re