免费在 GitHub 公开

baseline-repro

Internal Harness instruction source for baseline-repro. Route through visible Harness aliases instead of invoking directly.

中文简介正在准备中。当前显示原文说明。

实测评分未实测

包含的文件（3）

安装前可以查看交给 AI 代理的指令内容。

Read these first:

Use this skill for WF5 when the user wants baselines reproduced fairly before new method implementation.

Read the baseline candidates from docs/Refined_Idea.md, docs/30_evidence/Baseline_Table.md, and project state. For legacy projects, fall back to docs/Technical_Spec.md when no refined idea exists.
Create or refresh the first runnable project environment for the baselines and sync the ## Environment section in CLAUDE.md.
Use docs/Dataset_Stats.md and project context to align data and evaluation conditions.
Resolve baseline inclusion/skipping rules from the approved Baseline Contract when present, and resolve the canonical evaluation protocol from the approved Evaluation Contract when present; otherwise derive drafts from reproduced baseline evidence and persist the tracked metric names for WF10.
Reproduce each requested baseline with minimal environment-specific changes.
Compare reproduced metrics against paper-reported metrics.
Record baseline run artifact bundle paths for reproduced metrics.
Write docs/Baseline_Report.md using the canonical template.
Create or refresh docs/30_evidence/Baseline_Table.md as the human-readable Conclusion Evidence table for baseline repos, papers, configs, reproduced metrics, logs, and skip reasons.
If docs/20_facts/Codebase_Map.md exists, keep it synchronized when baseline directories, scripts, configs, or durable entry points change.
Update:

PROJECT_STATE.json baseline metrics
PROJECT_STATE.json evaluation protocol or tracked metrics for later WF10 comparison
project_map.json baseline status and entry point
CLAUDE.md environment facts and baseline reference

Keep adaptation notes, training config notes, and reproduced-versus-paper comparison.
Include evidence sources for reproduced metrics, commit hashes, configs, and any paper-reported values.
Include run artifact bundle paths for reproduced metrics.
Keep docs/30_evidence/Baseline_Table.md source-artifact oriented; it is human-readable Conclusion Evidence, not an .evidence/** Evidence Chain.
If no approved Baseline or Evaluation Contract exists, write the affected baseline/evaluation protocol as draft/derived and surface the approval gap.
WF5 is the first hard approval point for baseline/evaluation contracts in dynamic-context projects. Use protocol drift, context gates, docchain gates, and review packets when approval or explicit draft acceptance is needed.
Treat environment creation here as part of the canonical WF5 gate, not as a separate pre-workflow step.
Use the canonical pre-training commit rule for baseline code changes.
Report a Gate ledger for baseline reproduction, contract readiness, protocol drift, dynamic-context gates, docchain checks, workflow-state checks, and semantic commits. Mark any skipped gate NOT_RUN with the reason.
Treat template wording as structure-only; localize headings and narrative text according to ../../../.agents/references/language-policy.md unless a field is explicitly English-only.

Treat natural-language requests as the canonical $baseline-repro flow.
Preserve the original expectations around faithful reproduction and minimal baseline edits.
Use the Codex toolchain, but keep the canonical output files and state updates.

Follow the local prompt, baseline report template, and language policy instead of simplifying the reproduction stage.

temporary drafts.

还没有评价。欢迎分享你的使用感受。

无需登录即可发表。欢迎分享使用感受、提问，或说明在哪些环境下无法运行。

还没有评论。来写第一条吧。