Autoresearch: Autonomous Improvement Loop

Run a self-directed improvement loop on any measurable codebase metric. Inspired by Andrej Karpathy's approach to systematic, measurable iteration.

Core Protocol

For each iteration:
  1. REVIEW  — Read git log + results log (what worked/failed/untried)
  2. PICK    — Choose ONE focused change (highest expected impact)
  3. MAKE    — Implement the change (single file if possible)
  4. TEST    — Run test suite (must pass or revert)
  5. COMMIT  — Git commit with "experiment:" prefix BEFORE verification
  6. VERIFY  — Run mechanical benchmark/metric
  7. DECIDE  — If improved → keep. If worse → git revert
  8. LOG     — Append result to TSV log
  9. REPEAT  — Go to step 1