self-verify — independently scanned and version-tracked by SaferSkills.
SaferSkills independently audited self-verify (Agent Skill) and scored it 100/100 (green). The audit ran 55 deterministic rules across Security, Supply Chain, Maintenance, Transparency, and Community; it found 0 high-severity and 0 lower-severity findings. The full rule-by-rule trace and per-finding evidence are below. Free, methodology-open.
Findings & checks · 0 flagged
Every scanned point with the score it earned and what moved between them.
First recorded scan — no prior version to compare against.
The primary manifest — the file an agent reads to learn what this artifact does.
写作过程中的引用快速抽查。验证 Claude 输出的文献引用是否指向真实存在的 PubMed 文献。
不是什么:
与 ref-verifier 的关系:ref-verifier 是论文完稿后的系统性体检(全量 bib、逐条 claim-摘要比对、生成报告)。self-verify 是写作中的随身血压计(几条引用、几秒出结果)。
仅显式触发:
verify 一下 / self-verify / self-check核查一下 / 查一下刚才的引用 / check citations半自动提示:当 Claude 输出的内容包含 ≥2 条来自训练知识(非本次对话中 API/工具调用返回)的文献引用时,在输出末尾追加一行提示:"本段包含 N 条来自训练知识的引用,建议 verify 核查。"仅提示,不自动执行。
不触发的情况:引用来自本次对话中的 Entrez 搜索结果或用户提供的文献列表时,不提示。
能查的:
不查的(标 UNVERIFIABLE,附原因):
从目标文本中识别所有文献引用,提取结构化 anchor。
目标文本来源(按优先级):
提取规则:
Author et al. (YYYY)、Author YYYY、Author and Author (YYYY)、PMID XXXXXXXX、[N] 编号引用(如能从上下文推断对应文献){
"id": 1,
"citation_text": "Treece et al. (2010)",
"anchor": {
"author_surname": "Treece",
"year": "2010",
"title_keywords": "cortical bone mapping"
}
}title_keywords 从引用的上下文语境中推断(被引用时提到的方法名、主题词等),取 2-5 个实词提取质量自检:
对每条 anchor 调用验证脚本:
python skills/self-verify/scripts/entrez_verify.py --author <surname> --year <YYYY> --title-keywords "<kw1> <kw2> <kw3>"脚本为确定性 Python,不依赖 LLM 判断。输出 JSON:
{
"verdict": "PASS",
"pmid": "20176027",
"matched_fields": ["author", "year", "title_full"],
"evidence": "https://pubmed.ncbi.nlm.nih.gov/20176027/",
"suggested_fix": null
}四档判定:
| 档位 | 含义 | 条件 |
|---|---|---|
| PASS | 三重匹配全中 | author + year + 全部 title_keywords 命中同一 PMID |
| PARTIAL | 部分匹配 | author 对但年份偏差 ≤1,或标题关键词部分命中 |
| FAIL | 未找到匹配 | Entrez 无匹配,或匹配到的记录核心字段对不上 |
| UNVERIFIABLE | 无法核查 | anchor 不完整 / Entrez 不可达 / 非 PubMed 文献 |
关键约束:
输出 Markdown 表格:
## self-verify 报告
目标: <被核查文本的一句话描述>
引用总数: N | PASS X | PARTIAL Y | FAIL Z | UNVERIFIABLE W
| # | 引用 | verdict | PMID | evidence | suggested_fix |
|---|------|---------|------|----------|---------------|
| 1 | Treece et al. 2010 | PASS | 20176027 | [link] | — |
| 2 | Smith et al. 2023 | FAIL | — | — | 未找到匹配,检查作者拼写 |suggested_fix(基于 Entrez 实际返回的最近匹配)reason每次使用时不需要主动声明,但如果用户对结果有疑问,说明:
~30 seconds. Free. No account. Every finding cites a rule and a line of evidence.