autoresearchlisted

# Autoresearch Autonomous **hill-climbing** over a **mutable artifact**, judged by a **frozen referee**, advanced by a **git ratchet**: keep what measurably improves, revert what doesn't, never stop until the stopping condition fires. The human programs the protocol; you run the experiments; the referee decides. Generalized from [karpathy/autoresearch](https://github.com/karpathy/autoresearch). **Continue mode**: if invoked as `continue` (or the repo already has a `PROTOCOL.md` and `results.tsv` on an `autoresearch/*` branch), read both, skip to phase 4, and resume the loop under the existing protocol. ## 1. Qualify The pattern stands on four legs. Name each concretely before anything else: | Leg | Question | Example | |---|---|---| | Mutable artifact | Which files may the loop edit? | `src/translator/**` | | Quantitative metric | One number; which direction is better? | tests passed, higher | | Bounded eval | One command, fixed budget, prints the metric | `pnpm eval`, ~30s | | Keep/discard | Can a bad change be fully reverted? | git commit / reset | A missing metric or eval command is normal: you construct it in phase 3. But if the artifact isn't revertible or nothing is measurable even in principle, say so and stop. This problem does not hill-climb. ## 2. Align: grill the protocol out of the user Invoke the `grill-with-docs` skill on the proposed run (fall back to `grilling` if unavailable). The grilling must pin down, at minimum: - **Objective**: the single metric