Hill-climb a measurable artifact: try → measure → keep-or-revert → log, on a branch, with a trail. Full procedure: references/sops/experiment-loop.md. Existing examples + use-case cards: experiments/.
First, the autonomy gate
Before looping unsupervised, confirm all three:
- Objective, cheap, automatable metric (code computes it in seconds–minutes).
- Reversible sandbox (git branch / throwaway data /
.tmp). - No PHI, no confidential data, no external send, no prod deploy.
If any fails → human-in-the-loop: propose each change, get approval, no overnight run. Regulated/production outcomes feed the human-reviewed change process — never auto-deploy. This is CLAUDE.md's scale-caution-to-stakes rule.
Run it
- Pick or write the experiment's
CARD.md(objective, metric+direction, the one mutable surface, budget, autonomy, owner). New ones can copyexperiments/forecast-tuning/. git checkout -b experiments/<tag>.- Baseline run →
tools/experiment_log.py add <results.tsv> --id ... --metric ... --status keep --note baseline. - Loop: edit the one mutable surface → run the read-only harness → log → keep (commit) if improved, else
git reset. Simpler-and-equal = keep. - Stop on interrupt or the budget. Hand the winner + trail to the human.
Guardrails
- The eval harness is read-only ground truth — never edit it to game the metric.
- One mutable surface only (reviewable diffs).
- Run artifacts (
results.tsv,run.log,data/) are git-ignored; the harness + surface + CARD are the tracked blueprint. - Don't ship regulated/production changes from the loop.