AgentSkillsCN

cite-guard-ground

阶段依据：评估每一条引用是否与其所支持的主张相匹配；优先获取md、html、tex、rtf、txt、pdf等格式的证据；填写ground_*列，并生成依据报告。

SKILL.md
--- frontmatter
name: cite-guard-ground
description: >
  Stage ground: evaluates whether each reference supports the claims it is cited for; fetches evidence preferring md/html/tex/rtf/txt/pdf; writes ground_* columns and produces grounding reports.
license: MIT
metadata:
  version: 1.0.0
  category: research-qa
  tags: [grounding, claims, evidence]

cite-guard-ground

Extracts atomic claim-sentences from TeX, maps cited keys, then for each reference:

  • •fetches best available evidence (md > html > tex > rtf > txt > pdf)
  • •selects snippets with keyword overlap
  • •computes a heuristic support score (supported/weak/unsupported/contradicted)
  • •emits safe LaTeX rewrite suggestions for high-priority unsupported claims

Per reference outputs:

  • •ground_quality [0-100]
  • •ground_confidence [0-100]
  • •ground_remediation

Artifacts:

  • •out/claims.json
  • •out/grounding_report.md
  • •out/rewrites.tex
  • •out/evidence_cache/

Run: /cite_guard_ground or python3 ... ground --fetch