autoresearch

name	autoresearch
description	Karpathy-style keep/revert experiment loop for bounded targets with external metrics. Use with atris/experiments packs.
version	1.0.0

Autoresearch Skill

Use this when improving a bounded target through experiments instead of intuition.

Loop

1. Read program.md
2. Identify the one bounded mutation target
3. Run measure.py to get the baseline
4. Apply one candidate change
5. Run the same metric again
6. Keep only if the score improves
7. Append the result to results.tsv
8. Revert regressions immediately

Hard rules

• One target per experiment pack
• External metric only; no self-scored wins
• Keep/revert must be deterministic
• results.tsv stays append-only
• If the metric is noisy, document the margin before keeping

In this repo

Run experiments from atris/experiments/.

Good first lanes:

• worker-routing
• claim conflict behavior
• summary quality
• CLI ergonomics with fixed replay cases

Commands

atris experiments validate
atris experiments benchmark
atris experiments init <slug>

Avoid

• changing multiple surfaces at once
• vague goals like "make it better"
• keeping a change without a measured improvement