| name | research-autoresearch-loop |
| description | Use when running human-supervised iterative research experiments, recording autoresearch results, maintaining resume state, applying verify/guard gates, or coordinating local_mac smoke tests with remote_desktop_4060 formal runs. |
Research Autoresearch Loop
Use this skill for human-AI collaborative experiment iteration. It borrows useful ideas from autoresearch systems but keeps the user in control of research direction, claims, and final decisions.
Core Rules
- Do not run unattended autonomous research.
- Every iteration must target a claim, hypothesis, metric, or integrity risk.
- Record iterations in
docs/thesis/autoresearch-results.tsv.
- Record resumable state in
docs/thesis/autoresearch-state.json.
- For formal runs, generate a lightweight report in
docs/thesis/experiment-reports/EXP-*.md with baseline comparison, verify/guard status, and environment snapshot status.
- Use a dual gate: verify improvement, then guard against invalid science.
- Formal GPU work defaults to
remote_desktop_4060; cloud_autodl is fallback.
Workflow
Read references/loop.md for the iteration contract. Read references/source-map.md for provenance.
- Choose the target claim, current best run, primary metric, and candidate change.
- Confirm code quality and experiment contract before remote GPU work.
- Run
local_mac smoke test.
- Run formal experiment on
remote_desktop_4060 when needed.
- Recover artifacts and update
experiment-registry.md.
- Apply verify gate: did the experiment answer the question or improve the metric?
- Apply guard gate: no leakage, config drift, phantom result, or claim inflation.
- Record the iteration with
scripts/new_autoresearch_iteration.py.
- Generate or update the experiment report with
scripts/new_experiment_report.py --experiment-id EXP-... --baseline EXP-... when a baseline comparison is relevant.
- Hand reviewed results to
$research-results-analysis.
Output Contract
Always include:
- target claim or research question
- current best run
- candidate experiment and expected improvement
- local smoke-test command
- remote target and recovery path
- verify gate and guard gate
- registry and autoresearch files to update
- experiment report and baseline comparison status
- human decision needed before promotion
