| 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
- Read
program.md
- Identify the one bounded mutation target
- Run
measure.py to get the baseline
- Apply one candidate change
- Run the same metric again
- Keep only if the score improves
- Append the result to
results.tsv
- 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
