Cite-Bound Content Generator (Phase 4 Turn 1)

LLM Prompt Constraints (enforced inside ask_llm callback)

"You write report HTML for the disease/region described in the contract. Every fact claim (drug name, brand name, dosage, number, recommendation grade, epidemiology, mechanism) MUST carry a citation anchor in one of these forms: [guideline:<source_id>:<locator>] / [pmid:<id>:<locator>] / [nct:<id>:<locator>]. The source_id of every anchor MUST appear in the provided sources_summary. Sentences without an anchor for a fact claim are forbidden — your report will be rejected."

Agent-Authored Compose Artifact Contract

Production plugin runs should pass report content through .cache/<slug>/agent_authored_compose.json instead of letting a callback write HTML/model/flow content. The callback may read this JSON and return it to compose(), but the content remains subject to all compose and verification gates.

Required JSON keys:

• html: report HTML with citation anchors; direct .html files are not the primary contract
• claims: citation index candidates using the same claim_text / anchor / position / classification shape
• provenance: must include authoring_mode=agent_authored_compose_artifact, generated_by=codex_or_claude_code_plugin_executor, and current source_run_slug
• source_usage: must show body utilization for PubMed, ClinicalTrials.gov/AACT, and PubTator/entity channels when those sources were fetched
• formula_dag_ref: path/reference to reusable model output; formula DAG is not authored by the callback
• mermaid_sources: decision-tree sources to be rendered and audited by mermaid-renderer

compose() persists compose_authoring_provenance.json and mirrors authoring_provenance into citations_index.json. Production smoke must block if this provenance is absent or says callback-authored content.

Final Compose Artifact Gate

Production plugin continuation must reject precompose artifacts. .cache/<slug>/agent_authored_compose.json is acceptable only when provenance.phase=final, provenance.source_run_slug matches the current slug, reader_layer.html is non-empty, and audit_layer.evidence_support_table plus audit_layer.reader_to_audit_mapping are non-empty. A precompose checkpoint, source index, or process stub cannot be promoted to report body.

IFI 章节强制结构(每次 compose 必读 — R1 by Phase-2-Quality-Fix)

参考完整模板:references/ifi-section-template.md.

输出 html 必须含以下结构 (composer.py _assert_ifi_structure 在 enforce_ifi=True 时强制,违者抛 ComposeError 含所有缺失项 + orchestrator retry loop 反馈给 LLM 一次性补齐):

• 7 个 <section data-section="..."> marker:exec-summary / epidemiology / treatment-landscape / market-sizing / competitive-dynamics / lp-framework / appendix
• market-sizing 内三 <section data-subsection="tam|sam|som"> 子段
• <div class="mermaid">...</div> 决策树 block ≥ 4 张
• LP 提及次数 ≥ 50；LP1 / LP 1 / LP2 等编号写法必须计入，但不能只靠关键词堆砌通过
• 主报告 text 长度 ≥ 4000 字(去 html tag)
• H2/H3 深度 ≥ 10, Mermaid 每张必须有实际路径逻辑(不是 A-->B toy diagram)
• Mermaid 必须包含可审计的交叉点/分流/LP 数字；数字必须带 citation。没有 cohort × pathway × LP 交叉数字的流程图不能算 V25 级图。
• Mermaid 中每个数字标签必须由就近 locator anchor 支撑；不能把 PCP>4.5%、2.4 per 1000、24小时 等数字锚到不含该数字的 PMID 或泛化来源。
• 所有市场/统计数字必须锚到含该数字或公式证据的源文件；派生数字必须先写入可引用的 model/evidence 文件后再引用。
• citation anchor 必须带 locator，例如 [guideline:ID:line:42] / [pmid:123:abstract] / [evidence:path:line:12]；无 locator anchor 不能支撑生产数字 claim。
• 正文必须实际使用本轮多渠道来源，不能只反复引用 guideline 或内部 model 文件；PubMed/PMC/NCT/BioC/entity/evidence-grading 产物进入报告论证才算进入链路。来源索引或附录清单不能单独满足正文来源利用率。
• ClinicalTrials.gov/AACT 与 PubTator/entity recall 如果已经保存到本轮 sources/，正文必须用 [nct:...] / [aact:...] / [pubtator:...] 锚点进入论证；只抓取、只列索引、只写进附录都不算进入任务链。
• medical-evidence-grading 标为 D 的 PubMed 记录只能做背景或附录参考，不能作为正文结论、市场数字、LP 建议或药物-病原判断的支撑锚点。
• 公式 DAG / model evidence 只能支撑派生数字，不能成为正文主要 citation 来源；正文必须主要锚到指南、文献、试验、全文、实体证据和本轮保存的模型证据组合。
• 禁止把“证据消化矩阵”反复作为正文主体。证据矩阵只能帮助组织材料，最终报告必须写成可读的 V25 级市场叙事、队列分析、路径判断和 LP 策略。
• 文献召回 >50 是硬门槛而非停止点；呼吸科 IFI 等热门领域应以 >100 saved literature sources 为正常目标，并在正文分析中呈现主题化利用。
• 正文必须包含核心证据综合，而不是只把 PubMed/PMC/试验堆进附录。对于 >100 PubMed saved sources 的生产报告，非附录正文必须有足够 unique PMID 锚点进入诊断、队列、病原、治疗、预后或 LP 方法论分析。
• V25 对标不仅是标题/表格/图片数量；正文应包含增长驱动、支付/医保/准入/院内采购影响、LP排序方法，以及方法/证据等级/敏感性说明。缺少这些战略深度模块的报告不能算 V25 级。
• 禁止模板堆砌：不要用 Block 1..N、同句复制、批量替换队列名、重复“执行含义/质量要求”段落来凑字数、表格数或 LP 次数。
• 禁止机械审计口吻：不要反复输出 来源见本行、公式输入核对项、本行证据、证据项、本项仅按、队列动作N/模块N 等检测痕迹。证据说明必须融入自然市场报告 prose 或放入独立来源索引；agent_authored_compose.json.claims[*].claim_text 同样不能残留这些短语。
• 禁止把 source inventory / 已保存于 sources / 附录证据说明 / 待 worker / 后续量化模型 当正文
• 每数字 + 每事实声明带 citation 锚点(P0 铁律,无锚点 = 凭记忆 = 阻断)

sub-page 输出指令(R3b by Phase-2-Quality-Fix · 每 cohort 独立 .html)

每 cohort(基于 staging.dimensions 的 sub_cohorts)用 <sub-page slug="cohort-slug">...</sub-page> 包裹独立段:

<sub-page slug="1l-treatment">
  <section class="cohort-page" data-cohort="1l-treatment">
    <h1>1L 治疗 cohort 深度分析</h1>
    ...
  </section>
</sub-page>

composer._parse_sub_pages 解析后, ComposedReport.sub_pages: list[dict {slug, html}] 经 _persist_sub_pages 落盘成 output/<disease-slug>/page_<cohort-slug>.html,主报告通过 <a href="page_<slug>.html"> 链接。子页必须是实质 cohort 页:包含诊疗路径、市场测算/交叉点数字、表格、LP 动作和 citation；几百字节空壳页必须被下游 audit 阻断。

Functions

• compose(slug_dir, ask_llm, previous_violations=None, *, enforce_ifi=False, output_dir=None, return_report=False) -> (html_path, idx_path) | ComposedReport

  • enforce_ifi=True (production) → 强制 IFI 结构,违者 ComposeError
  • enforce_ifi=False (default, fake-LLM tests) → 只查 anchor 解析,不查结构
  • production orchestration must call return_report=True so sub_pages/toc_anchors are passed to bundle; losing ComposedReport is a workflow bug.

• _parse_sub_pages(raw_html) -> (main_html, sub_pages) — 抽 <sub-page slug="..."> 包裹的子页

• _extract_toc_anchors(html) -> list[str] — 从 main_html H2/H3 id 抽 TOC 浮动锚点

• ComposedReport(main_html, sub_pages, toc_anchors, claims, raw_html) — 升级输出 schema dataclass

P0 Watchdog

tests/test_cite_bound_content_generator.py::test_compose_p0_watchdog — grep skill source for forbidden hardcoded patterns. CI blocks on hit.