AGENT SKILL
autosota-research-loop
Use when operating research-os from a natural-language research goal through evidence-backed selection, redline-audited reproduction, parallel branch optimization, and skill distillation.
- License 许可
- MIT
- 最近更新
- 2026-06-19
- Stars
- ★ 1
- Forks
- 0
01 前言 Front Matter
- name
- autosota-research-loop
- description
- Use when operating research-os from a natural-language research goal through evidence-backed selection, redline-audited reproduction, parallel branch optimization, and skill distillation.
03 SKILL.md 正文
AutoSOTA Research Loop
Use this skill when the task is to turn a research request into a controlled research-os execution loop instead of an ad hoc paper chase or one-off experiment.
When To Use
- The user starts from a natural-language research objective.
- The task needs paper, code, dataset, and benchmark selection backed by recorded evidence.
- The work must recover a comparable baseline before optimization.
- The plan needs parallel branch exploration under explicit redline guardrails.
Required Operator Sequence
- Start with
research-os intake "<goal>"to convert the free-form request into a structured brief. - Build the evidence ledger and use it for evidence-backed selection of papers, repos, datasets, and benchmark constraints.
- Create the bundle from the intake record only after the sources, feasibility, and reproducibility signals are attached.
- Recover the baseline before proposing branch modifications or claiming any optimization path is valid.
- Run a redline audit on the reproduction plan and again on every optimization candidate.
- Generate parallel branches from the cleared baseline, preserving parent branch, idea source, patch state, run configuration, and keep-or-reject status.
- Rank branch outcomes by metric, stability, cost, and comparability confidence before selecting survivors.
- Distill repeated wins and stable repair patterns into notes or follow-on skills only after the evidence is recorded.
Guardrails
- Never skip
research-os intake; the loop starts from structured intake, not from improvised commands. - Never treat secondary summaries as enough when a primary paper, repo, dataset page, or benchmark source is available.
- Never optimize before a baseline exists and is reproducible on the selected runner.
- Never allow a branch to proceed without redline review of metric, dataset, split, script, and constraint integrity.
- Never report gains from non-comparable settings as if they were baseline-equivalent improvements.
- Never let parallel branches explode without a branch budget and survivor selection rule.
Branch Discipline
- Parallel branches are search instruments, not excuses to run untracked experiments.
- Each branch should represent one bounded idea family or a clearly stated transfer hypothesis.
- Reject branches that are expensive, fragile, or redline-unsafe even if they show short-term metric movement.