paper-claim-audit

name	paper-claim-audit
description	Audit manuscript numeric and strength-of-claim statements against a normalized ResearchOS evidence pack and result-to-claim mapping. Use before paper polishing, revision, compilation, or submission when the user needs a deterministic evidence-boundary report.
tools	["read_file","write_file","audit_paper_claims","finish_task"]
strict_tools	true
model_tier	standard
temperature	0.1
allowed_read_prefixes	["user_inputs/paper-claim-audit/","drafts/","_runtime/skill_sessions/"]
allowed_write_prefixes	["drafts/"]
outputs_expected	{"audit_markdown":"drafts/paper_claim_audit.md","audit_json":"drafts/paper_claim_audit.json"}
interaction	{"mode":"guided","language":"zh-CN","summary":"用证据包和 claim mapping 审计论文中的数值与强断言；失败会被保留，不会自动粉饰为通过。","request_required":false,"request_prompt":"可选：说明希望优先检查的章节或声明类型。","example_request":"重点检查摘要和实验部分是否有超出证据的提升幅度。","required_inputs":[{"id":"manuscript","label":"论文 LaTeX 稿件","description":"需要审计的完整稿件；若放在 user_inputs，先复制到标准 drafts 路径再运行工具。","paths":"[Truncated]","extensions":"[Truncated]","min_bytes":80,"example":"drafts/paper.tex"},{"id":"evidence_pack","label":"实验证据包","description":"标准化的可用指标和 mock 状态；没有它不能判断数值声明。","paths":"[Truncated]","extensions":"[Truncated]","min_bytes":20,"example":"drafts/experiment_evidence_pack.json"},{"id":"result_mapping","label":"结果到声明映射","description":"列出允许和禁止的措辞，防止把弱/模拟结果写成强结论。","paths":"[Truncated]","extensions":"[Truncated]","min_bytes":20,"example":"drafts/result_to_claim.json"}],"optional_inputs":[],"outputs":[{"id":"audit_markdown","label":"声明审计报告","path":"drafts/paper_claim_audit.md","description":"面向人阅读的失败、警告和证据边界。"},{"id":"audit_json","label":"结构化审计结果","path":"drafts/paper_claim_audit.json","description":"可供后续修订 Skill 和自动化检查读取。"}]}

Evidence Claim Audit

If a verified input was supplied under user_inputs/paper-claim-audit/, copy its exact content to the matching standard drafts/ path before invoking the audit. Record that staging action in the final summary; do not modify its semantics. Read drafts/writing_style.json and drafts/writing_storyline.md when present so the final summary can identify which venue-specific contribution chain is affected; their presence never changes the deterministic evidence verdict.

If a necessary material is absent after deterministic readiness passes, write user_inputs/paper-claim-audit/_followup_request.md with the exact required evidence path and call ask_human; do not reinterpret an unsupported result as a pass.

Run audit_paper_claims with the standard three drafts/ paths. Read both generated outputs. Treat every FAIL, mock-only marker, unsupported strong claim, or forbidden wording violation as a real failure. Summarize only what the deterministic audit found, point to the exact report paths, and finish. Do not edit the paper in this Skill; use paper-polish or paper-revision after the user reviews the report.