agentsclimarketplace

Paper digger evaluate

Skill Iu1ky/paper-digger/plugins/paper-digger/skills/paper-digger-evaluate

Use when issuing a research go/revise/pivot/stop verdict from value, novelty, rigor, evidence integrity, and anti-fabrication checks. Triggers on 价值评估, 诚信稽核, research red-team, go/no-go.From its SKILL.md

Install
npx -y skills add Iu1ky/paper-digger --skill paper-digger-evaluate

Assembled from the repository path, not quoted from the project. Check it against their README if it does not work.

2 things to look at

  • 16 days oldThe repository was created 16 days ago. New is not bad, but a brand new repository carrying a familiar-sounding name is the shape a typosquat arrives in, and there has been no time for anyone else to find a problem with it.
  • 0 stars0 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.

SKILL.md

3.0 KB, ~1.1k tokens by cl100k_base, as published. Nobody here has run it

Paper Digger Evaluate — 研究期价值与诚信红队

在 4 个节点对「当前工作」做对抗式、多维度评价,回答:这值得做/做得够好吗?我们是不是在自欺/造假? —— 输出可执行 verdict。区别于 Phase 8 的投稿期外部审稿模拟,这是研究期内部红队

运行时

定位同级 paper-digger skill。调用 paper_digger.* API 前,把它的 scripts/ 目录加入 Python sys.pathPYTHONPATH;不要依赖全局安装。向 lens 只提供 claim、指标摘要、路径、哈希和必要片段,不内联原始日志或大数组。

两轴 rubric

Axis A — 价值与水平

  • A1 价值/意义("so what?",抓无意义的工作)· A2 新颖性(抓低水平/增量)· A3 严谨与逻辑(抓逻辑问题)· A4 水平/雄心 · A5 证据充分性(随节点)

Axis B — 诚信与认知失效

  • B1 数据伪造 · B2 方法虚构 · B3 引用幻觉 · B4 思维固定(主动生成 ≥2 个反框架/竞争解释,检验 tunnel vision)

4 个节点(逐节点换侧重,见 NODE_FOCUS)

  • 节点 ① idea 确定:重 A1 价值 / A2 新颖 / B4 思维固定 → 立项 / 改框架 / 换题
  • 节点 ② 最小验证完成:重 A5 证据 / B1 伪造 / B2 方法虚构 → 继续 / 修正 / kill-early
  • 节点 ③ 初步成果:重 A3 逻辑 / A4 水平 / B1 / B3 → 继续 / 补实验 / 重定位
  • 节点 ④ 最终成果:全轴审计 + "so what?" 终检 → 进入写作 / 返工 / 放弃

执行流程

  1. 读取节点的 NODE_FOCUS 维度,确定本轮重点。
  2. 按 effort 运行对抗式 lens:
    • lean:先跑 1 个综合 lens;节点②/④或出现诚信信号时至少再跑 1 个独立 integrity lens。
    • standard:默认 2 个独立 lens(价值/证据、诚信/反框架)。
    • deep:仅在节点④、verdict 冲突、诚信红旗或用户明确要求时扩展到 novelty、rigor、integrity、fixed-thinking 4 个 lens。 每个 lens 使用 fresh context、默认怀疑并产出 {lens, axis: A|B, verdict: GREEN|YELLOW|RED, must_fix: [...], rationale}
  3. 汇总并落盘:用 paper_digger.evaluate.record(workspace, node, lens_verdicts, now=...) —— 它做 aggregate(worst-of + 任一 Axis-B RED 置 blocking_integrity),写 08_evaluation/eval_node<N>.md,并把 verdict 追加进 state.jsonevaluations[]
  4. 在 checkpoint 呈现 verdict + 必修项。RED 不自动 kill,由人拍板(GREEN 通过 / YELLOW 带必修项继续 / RED 强烈建议 pivot 或 kill)。

铁律

  • 评价 lens 必须独立、对抗,不得附和「我们很看好」的框架(避免自评盲点)。
  • 任一 Axis-B(诚信)RED 即 blocking_integrity,必须在 checkpoint 显著标出。
  • 结论只依据真实证据(evidence/ + verification);不臆测、不放水。

What ships with it: 1 file

257 B alongside SKILL.md

agents/

Keep looking

Skills are one crate of 326,696. Ordering is by how many stacks a row turns up in, so the top of any crate is what has actually been picked rather than what has the most stars.