dual-ai-paper-coach: Command for Claude Code

/claim-check — 双 AI Claim 校准

Claude 和 Codex 独立给论文 1-3 个核心 claim 打分，分歧 ≥ 20 的就是论文最脆的地方。

⚠️ 打分标杆：Science / Nature——Codex 按 S/N 审稿人的怀疑度打 evidence；S/N 标杆下 evidence ≥ 80 才算"经得起多学科读者审视"。Codex 打分偏低是设计上的特性。

步骤

步骤 1：提取 claims

读 my-paper/draft.md 的 Abstract，把 1-3 句"主张"提取成 YAML：

# my-paper/claims.yaml
claims:
  - id: C1
    text: "Global AOD declined 10% over 2010-2024"
    section: Results §3.1
    evidence_keys: [global_mean.trend_per_decade]
  - id: C2
    text: "Decline driven by East Asia anthropogenic reduction"
    section: Discussion §4.2
    evidence_keys: [regional_trends.east_asia]
  - id: C3
    text: "Aerosol policy has improved global air quality"
    section: Conclusions §5
    evidence_keys: []   # 推论性 claim，最容易被打低分

如果 claims.yaml 已存在，直接用现有的；否则你（Claude）提取并写入。

步骤 2：Claude 自评（你来做）

读 prompts/writer_system.md 但不要读 reviewer_system.md。基于 draft + analysis_results.json，给每个 claim 打两个维度：

• evidence（0-100）：数据支持度
• wording（0-100）：当前措辞强度（>80 = overclaim）

输出 JSON 到 my-paper/runs/<ts>/claim_score_claude.json：

{
  "C1": {"evidence": 85, "wording": 80, "reason_evidence": "...", "reason_wording": "..."},
  "C2": {"evidence": 75, "wording": 85, "reason_evidence": "...", "reason_wording": "..."},
  "C3": {"evidence": 60, "wording": 70, "reason_evidence": "...", "reason_wording": "..."}
}

步骤 3：Codex 自评（通过 Bash 触发）

codex exec --skip-git-repo-check --sandbox workspace-write \
  -c 'model_reasoning_effort="xhigh"' \
  "Read AGENTS.md, my-paper/draft.md, my-paper/claims.yaml, and my-paper/data/analysis_results.json. Apply Science/Nature reviewer standards: a multidisciplinary PhD reader must find the evidence convincing. For each claim, score evidence (0-100; how strongly the data + reported statistics support the claim at S/N standard) and wording (0-100; >80 = overclaim that S/N reviewers would reject). Output JSON to my-paper/runs/<ts>/claim_score_codex.json."

关键参数（与 /codex-review 同源）：

• --sandbox workspace-write 必须——Codex 默认沙盒只读，写不进 JSON
• -c 'model_reasoning_effort="xhigh"' 审稿用 xhigh，默认 medium 不够严