@hongmaple0820/med-scale-research-os 0.43.0

package/target-research-platform/src/output/manuscript_quality.py

@@ -0,0 +1,401 @@
+"""
+Quality gate for generated manuscript deliverables.
+
+The gate is intentionally evidence-based: it checks files that were actually
+written to disk and flags unverifiable submission claims before the draft is
+shown as delivery-ready.
+"""
+from __future__ import annotations
+
+import json
+from pathlib import Path
+from typing import Any
+
+from .usage_envelope import collect_usage_envelopes, summarize_usage_envelopes
+
+
+REQUIRED_MARKDOWN_SECTIONS = (
+    "## 摘要",
+    "## 1 引言",
+    "## 2 材料与方法",
+    "## 3 结果",
+    "### 3.6.1 逐篇文献深度解读",
+    "## 4 讨论",
+    "## 5 结论",
+    "## 参考文献",
+)
+
+PROHIBITED_MARKDOWN_PHRASES = (
+    "LLM：None",
+    "LLM: None",
+)
+
+PROHIBITED_JOURNAL_CLAIMS = (
+    "国家自然科学基金",
+    "不存在利益冲突",
+    "免予伦理审查",
+)
+
+PROHIBITED_SUBMISSION_READY_CLAIMS = (
+    "可直接投稿",
+    "投稿级终稿",
+    "无需人工审校",
+    "无需专家审稿",
+    "submission-ready",
+)
+
+
+def _read_text(path: str | None) -> str:
+    if not path:
+        return ""
+    try:
+        return Path(path).read_text(encoding="utf-8")
+    except Exception:
+        return ""
+
+
+def _path_exists(path: str | None) -> bool:
+    return bool(path) and Path(path).exists()
+
+
+def _file_nonempty(path: str | None, min_bytes: int = 1) -> bool:
+    return bool(path) and Path(path).is_file() and Path(path).stat().st_size >= min_bytes
+
+
+def _read_json(path: str | None) -> dict[str, Any]:
+    if not path or not Path(path).is_file():
+        return {}
+    try:
+        payload = json.loads(Path(path).read_text(encoding="utf-8"))
+        return payload if isinstance(payload, dict) else {}
+    except Exception:
+        return {}
+
+
+def _add_check(
+    checks: list[dict[str, Any]],
+    check_id: str,
+    passed: bool,
+    *,
+    severity: str = "error",
+    message: str,
+    evidence: Any = None,
+) -> None:
+    checks.append(
+        {
+            "id": check_id,
+            "passed": bool(passed),
+            "severity": severity,
+            "message": message,
+            "evidence": evidence,
+        }
+    )
+
+
+def validate_manuscript_delivery(result: dict[str, Any]) -> dict[str, Any]:
+    """Validate a generated manuscript result and return a machine-readable gate."""
+    paths = result.get("paths") or {}
+    markdown_path = paths.get("markdown")
+    zhongliu_path = paths.get("zhongliu")
+    metadata_path = paths.get("metadata")
+    fact_check_path = paths.get("fact_check")
+    delivery_manifest_path = paths.get("delivery_manifest")
+    docx_path = paths.get("docx")
+    figures_dir = paths.get("figures_dir")
+    figures = result.get("figures") or []
+    reviews = result.get("literature_reviews") or {}
+
+    markdown = _read_text(markdown_path)
+    zhongliu = _read_text(zhongliu_path)
+    metadata = _read_json(metadata_path)
+    fact_check = result.get("fact_check") if isinstance(result.get("fact_check"), dict) else _read_json(fact_check_path)
+    delivery_manifest = _read_json(delivery_manifest_path)
+    usage_envelopes = collect_usage_envelopes(metadata)
+    checks: list[dict[str, Any]] = []
+
+    _add_check(
+        checks,
+        "result.success",
+        bool(result.get("success")),
+        message="Manuscript generator reported success.",
+        evidence=result.get("success"),
+    )
+    _add_check(
+        checks,
+        "artifact.markdown_exists",
+        _file_nonempty(markdown_path, 1024),
+        message="Chinese IMRaD markdown exists and is non-empty.",
+        evidence=markdown_path,
+    )
+    _add_check(
+        checks,
+        "artifact.zhongliu_exists",
+        _file_nonempty(zhongliu_path, 1024),
+        message="Zhonghua Oncology formatted markdown exists and is non-empty.",
+        evidence=zhongliu_path,
+    )
+    _add_check(
+        checks,
+        "artifact.metadata_exists",
+        _file_nonempty(metadata_path, 16),
+        message="Manuscript metadata JSON exists.",
+        evidence=metadata_path,
+    )
+    _add_check(
+        checks,
+        "artifact.fact_check_exists",
+        _file_nonempty(fact_check_path, 16),
+        message="Manuscript fact-check report exists.",
+        evidence=fact_check_path,
+    )
+    _add_check(
+        checks,
+        "delivery.manifest_exists",
+        _file_nonempty(delivery_manifest_path, 16),
+        message="Delivery manifest exists for the researcher package.",
+        evidence=delivery_manifest_path,
+    )
+    _add_check(
+        checks,
+        "artifact.docx_exists",
+        _file_nonempty(docx_path, 1024),
+        severity="warning",
+        message="Word draft exists; warning only because DOCX rendering is optional in minimal environments.",
+        evidence=docx_path,
+    )
+    _add_check(
+        checks,
+        "artifact.figures_dir_exists",
+        _path_exists(figures_dir) and Path(figures_dir).is_dir(),
+        message="Figures directory exists.",
+        evidence=figures_dir,
+    )
+
+    missing_sections = [section for section in REQUIRED_MARKDOWN_SECTIONS if section not in markdown]
+    _add_check(
+        checks,
+        "structure.required_sections",
+        not missing_sections,
+        message="Markdown contains the required journal manuscript sections.",
+        evidence={"missing": missing_sections},
+    )
+
+    char_count = int(result.get("char_count") or len(markdown))
+    _add_check(
+        checks,
+        "content.minimum_length",
+        char_count >= 3000,
+        message="Manuscript has enough body text for a substantive draft.",
+        evidence={"char_count": char_count, "minimum": 3000},
+    )
+    bad_markdown = [phrase for phrase in PROHIBITED_MARKDOWN_PHRASES if phrase in markdown]
+    _add_check(
+        checks,
+        "content.no_llm_none",
+        not bad_markdown,
+        message="Markdown must not expose internal placeholder values such as LLM: None.",
+        evidence={"found": bad_markdown},
+    )
+
+    bad_claims = [phrase for phrase in PROHIBITED_JOURNAL_CLAIMS if phrase in zhongliu]
+    _add_check(
+        checks,
+        "journal.no_unverified_claims",
+        not bad_claims,
+        message="Journal template must not assert unverified fund, conflict, or ethics claims.",
+        evidence={"found": bad_claims},
+    )
+    submission_ready_claims = [
+        phrase for phrase in PROHIBITED_SUBMISSION_READY_CLAIMS
+        if phrase.lower() in markdown.lower() or phrase.lower() in zhongliu.lower()
+    ]
+    _add_check(
+        checks,
+        "journal.no_submission_ready_claim",
+        not submission_ready_claims,
+        message="Draft must not claim it is submission-ready or expert-reviewed.",
+        evidence={"found": submission_ready_claims},
+    )
+    _add_check(
+        checks,
+        "journal.zhongliu_format",
+        "中华肿瘤杂志" in zhongliu and "【摘要】" in zhongliu,
+        message="Zhonghua Oncology wrapper contains journal marker and structured abstract marker.",
+        evidence=zhongliu_path,
+    )
+
+    review_count = int(reviews.get("count") or 0)
+    source_counts = reviews.get("source_counts") or {}
+    _add_check(
+        checks,
+        "literature.review_sources",
+        review_count > 0 and bool(source_counts),
+        message="Literature reviews include source accounting for traceability.",
+        evidence={"count": review_count, "source_counts": source_counts},
+    )
+    expert_like_reviews = (
+        int(source_counts.get("agent_cache") or 0)
+        + int(source_counts.get("openai") or 0)
+        + int(source_counts.get("scale_local") or 0)
+    )
+    expert_like_ratio = expert_like_reviews / review_count if review_count else 0
+    _add_check(
+        checks,
+        "literature.agent_or_llm_ratio",
+        review_count > 0 and expert_like_ratio >= 0.5,
+        severity="warning",
+        message="At least half of literature reviews should come from Agent cache or an LLM provider for Beta delivery.",
+        evidence={"ratio": round(expert_like_ratio, 3), "source_counts": source_counts},
+    )
+    llm_provider_reviews = int(source_counts.get("openai") or 0) + int(source_counts.get("scale_local") or 0)
+    _add_check(
+        checks,
+        "token_usage.envelope_present",
+        llm_provider_reviews == 0 or len(usage_envelopes) >= llm_provider_reviews,
+        severity="warning",
+        message="LLM-backed literature reviews should expose MedSCALE usage envelopes for token/cost accounting.",
+        evidence={"llm_provider_reviews": llm_provider_reviews, "usage_envelopes": len(usage_envelopes)},
+    )
+    references = metadata.get("references") if isinstance(metadata, dict) else []
+    reference_count = len(references) if isinstance(references, list) else int(result.get("reference_count") or 0)
+    _add_check(
+        checks,
+        "references.minimum_count",
+        reference_count >= 5,
+        severity="warning",
+        message="Researcher-facing manuscripts should include at least five traceable references.",
+        evidence={"reference_count": reference_count, "minimum": 5},
+    )
+
+    figure_files = []
+    missing_figures = []
+    for figure in figures:
+        path = figure.get("path")
+        if _file_nonempty(path, 128):
+            figure_files.append(path)
+        else:
+            missing_figures.append(path)
+    _add_check(
+        checks,
+        "figures.files_exist",
+        len(figure_files) == len(figures) and len(figures) > 0,
+        message="All registered figure files exist and are non-empty.",
+        evidence={"figures": len(figures), "missing": missing_figures},
+    )
+    figure_refs_missing = []
+    for idx, figure in enumerate(figures, start=1):
+        tokens = [f"图{idx}", f"图 {idx}", f"Figure {idx}", str(figure.get("relative") or "")]
+        if not any(token and token in markdown for token in tokens):
+            figure_refs_missing.append(figure.get("id") or idx)
+    _add_check(
+        checks,
+        "figures.referenced_in_text",
+        len(figures) > 0 and not figure_refs_missing,
+        severity="warning",
+        message="Registered figures should be referenced in the manuscript body.",
+        evidence={"missing": figure_refs_missing},
+    )
+    _add_check(
+        checks,
+        "figures.publication_set",
+        len(figures) >= 4,
+        severity="warning",
+        message="A full publication draft should include at least four manuscript figures.",
+        evidence={"figures": len(figures), "minimum": 4},
+    )
+
+    metadata_ok = bool(metadata.get("target") and metadata.get("disease"))
+    _add_check(
+        checks,
+        "metadata.readable",
+        metadata_ok,
+        message="Metadata JSON is readable and includes target/disease.",
+        evidence=metadata_path,
+    )
+    trace_sources = []
+    if isinstance(metadata.get("data_sources"), list):
+        trace_sources.extend(metadata.get("data_sources") or [])
+    enrichment = metadata.get("literature_enrichment") if isinstance(metadata, dict) else {}
+    if isinstance(enrichment, dict):
+        trace_sources.extend(enrichment.get("sources") or [])
+    has_literature_enrichment = False
+    if isinstance(enrichment, dict):
+        has_literature_enrichment = bool(enrichment.get("sources")) or any(
+            [
+                enrichment.get("pubmed_efetch"),
+                enrichment.get("openalex_used"),
+                enrichment.get("aminer_used"),
+                int(enrichment.get("mesh_count") or 0) > 0,
+                int(enrichment.get("citation_count_available") or 0) > 0,
+            ]
+        )
+    _add_check(
+        checks,
+        "metadata.source_traceability",
+        bool(trace_sources) and reference_count > 0,
+        message="Metadata includes data sources and references for traceability.",
+        evidence={"data_sources": trace_sources, "reference_count": reference_count},
+    )
+    _add_check(
+        checks,
+        "literature.enrichment_coverage",
+        has_literature_enrichment,
+        severity="warning",
+        message="Literature metadata should include PubMed MeSH or citation enrichment for Beta delivery.",
+        evidence=enrichment,
+    )
+    _add_check(
+        checks,
+        "fact_check.passed",
+        bool(fact_check.get("passed")),
+        message="Manuscript fact-check report passed without hard errors.",
+        evidence={
+            "score": fact_check.get("score"),
+            "errors": fact_check.get("errors"),
+            "warnings": fact_check.get("warnings"),
+        },
+    )
+    _add_check(
+        checks,
+        "delivery.manifest_files",
+        bool(delivery_manifest.get("files")),
+        severity="warning",
+        message="Delivery manifest should list package files with hashes.",
+        evidence={"file_count": len(delivery_manifest.get("files") or [])},
+    )
+    disclaimer = str(result.get("disclaimer") or "")
+    boundary_text = markdown + "\n" + zhongliu + "\n" + disclaimer
+    _add_check(
+        checks,
+        "review.boundary_statement_present",
+        any(token in boundary_text for token in ("投稿前", "专家", "伦理", "not a submission-ready", "does not replace")),
+        message="Draft includes visible boundaries requiring human review before submission.",
+        evidence={"disclaimer": bool(disclaimer)},
+    )
+
+    error_failures = [
+        check for check in checks
+        if check["severity"] == "error" and not check["passed"]
+    ]
+    warning_failures = [
+        check for check in checks
+        if check["severity"] == "warning" and not check["passed"]
+    ]
+    score = max(0, 100 - len(error_failures) * 20 - len(warning_failures) * 5)
+    if error_failures:
+        grade = "D"
+    elif len(warning_failures) >= 3:
+        grade = "C"
+    elif warning_failures:
+        grade = "B"
+    else:
+        grade = "A"
+    return {
+        "passed": len(error_failures) == 0,
+        "score": score,
+        "grade": grade,
+        "errors": len(error_failures),
+        "warnings": len(warning_failures),
+        "usage_summary": summarize_usage_envelopes(usage_envelopes),
+        "checks": checks,
+    }

package/target-research-platform/src/output/medical_skills_bridge.py

@@ -0,0 +1,140 @@
+"""
+医学研究 Agent Skills 桥接层
+
+将 AIPOCH / 循证医学工作流中的核心逻辑编码为可复用 Python 规则，
+供稿件生成管线在 PubMed 富化与 LLM 评述前调用。
+
+参考能力域（Evidence Insight）：
+- study-design-identifier：研究设计识别
+- paper-to-claim-verifier：主张可证性约束
+- drug-target-evidence-landscape：靶点证据景观 framing
+"""
+from __future__ import annotations
+
+import re
+from typing import Any
+
+
+# AIPOCH paper-to-claim-verifier 约束：禁止超出元数据的断言
+CLAIM_GUARDRAILS_ZH = """
+【证据边界】
+1. 仅根据已提供的标题、摘要、MeSH、发表类型、引用数评述，不得编造 p 值、HR、样本量。
+2. 摘要缺失时须标注「据标题推断」。
+3. 综述/Meta 分析不得表述为原创实验结论。
+4. 临床试验类文献须区分试验阶段与主要终点是否在给定摘要中出现。
+"""
+
+
+def _matches_marker(values: list[str], marker: str, *, ignore_topic_suffix: bool = False) -> bool:
+    for value in values:
+        current = value.lower()
+        if ignore_topic_suffix and current.endswith(" as topic"):
+            continue
+        if marker in current:
+            return True
+    return False
+
+
+def infer_study_design(paper: dict[str, Any]) -> str:
+    """
+    研究设计识别（对标 AIPOCH study-design-identifier）。
+
+    优先级：发表类型 > MeSH > 标题关键词。
+    """
+    pub_types = [str(x).lower() for x in (paper.get("publication_types") or [])]
+    mesh = [str(x).lower() for x in (paper.get("mesh_terms") or [])]
+    title = (paper.get("title") or "").lower()
+
+    pub_type_checks: list[tuple[str, list[str]]] = [
+        ("系统综述/Meta分析", ["meta-analysis", "systematic review", "network meta"]),
+        ("综述", ["review", "narrative review"]),
+        ("随机对照试验", ["randomized controlled trial", "randomised controlled trial"]),
+        ("临床试验", ["clinical trial", "clinical study"]),
+        ("队列研究", ["cohort studies", "prospective studies", "retrospective studies"]),
+        ("病例对照研究", ["case-control studies"]),
+        ("临床前/体外研究", ["preclinical", "in vitro", "xenograft", "cell line"]),
+        ("生物信息学/计算研究", ["bioinformatics", "transcriptom", "single-cell", "scrnaseq", "tcga"]),
+    ]
+    for label, keys in pub_type_checks:
+        if any(_matches_marker(pub_types, key) for key in keys):
+            return label
+
+    secondary_checks: list[tuple[str, list[str]]] = [
+        ("系统综述/Meta分析", ["meta-analysis", "systematic review", "network meta"]),
+        ("随机对照试验", ["randomized controlled trial", "randomised controlled trial"]),
+        ("临床试验", ["clinical trial", "clinical study"]),
+        ("队列研究", ["cohort studies", "prospective studies", "retrospective studies"]),
+        ("病例对照研究", ["case-control studies"]),
+        ("临床前/体外研究", ["preclinical", "in vitro", "xenograft", "cell line"]),
+        ("综述", ["review", "narrative review"]),
+        ("生物信息学/计算研究", ["bioinformatics", "transcriptom", "single-cell", "scrnaseq", "tcga"]),
+    ]
+    for label, keys in secondary_checks:
+        for key in keys:
+            if _matches_marker(mesh, key, ignore_topic_suffix=True):
+                return label
+            if key in title:
+                return label
+
+    if "review" in title:
+        return "综述"
+    if re.search(r"phase\s+[i123]", title):
+        return "临床试验"
+    return "机制与转化研究"
+
+
+def evidence_level_label(study_design: str, paper: dict[str, Any]) -> str:
+    """粗略证据等级（用于评述语气，非正式 GRADE）。"""
+    design = study_design.lower()
+    if "meta" in design or "系统综述" in design:
+        return "I（系统综述/Meta）"
+    if "随机" in design or "rct" in design:
+        return "II（RCT）"
+    if "临床试验" in design or "队列" in design:
+        return "III（临床/队列）"
+    if "综述" in design:
+        return "IV（叙述性综述）"
+    cites = paper.get("aminer_citation_count")
+    if isinstance(cites, (int, float)) and cites >= 50:
+        return "III（高被引机制研究）"
+    return "IV（临床前/探索性）"
+
+
+def build_literature_skill_context(paper: dict[str, Any], *, target: str, disease_zh: str) -> str:
+    """为 LLM 评述组装医学 skill 增强上下文块。"""
+    design = paper.get("study_design") or infer_study_design(paper)
+    level = evidence_level_label(design, paper)
+    mesh = "；".join((paper.get("mesh_terms") or [])[:8]) or "（无 MeSH）"
+    pub_types = "；".join((paper.get("publication_types") or [])[:5]) or "（未标注）"
+    keywords = "；".join((paper.get("keywords") or paper.get("aminer_keywords") or [])[:6]) or "（无）"
+
+    aminer_bits = []
+    if paper.get("aminer_citation_count") is not None:
+        aminer_bits.append(f"被引次数≈{paper['aminer_citation_count']}")
+    if paper.get("aminer_venue"):
+        aminer_bits.append(f"AMiner 期刊信号：{paper['aminer_venue']}")
+    if paper.get("aminer_cited_by_sample"):
+        sample = "；".join(paper["aminer_cited_by_sample"][:2])
+        aminer_bits.append(f"引用链样例：{sample}")
+
+    aminer_line = " | ".join(aminer_bits) if aminer_bits else "（未匹配 AMiner）"
+
+    return (
+        f"靶点/疾病：{target} / {disease_zh}\n"
+        f"研究设计（skill 推断）：{design}\n"
+        f"证据等级（粗估）：{level}\n"
+        f"MeSH：{mesh}\n"
+        f"发表类型：{pub_types}\n"
+        f"关键词：{keywords}\n"
+        f"AMiner 补充：{aminer_line}\n"
+        f"{CLAIM_GUARDRAILS_ZH}"
+    )
+
+
+def drug_target_evidence_framing(target: str, disease_zh: str, paper_count: int) -> str:
+    """靶点证据景观 framing（对标 drug-target-evidence-landscape）。"""
+    return (
+        f"在 {disease_zh} 语境下，{target} 相关公开文献约 {paper_count} 篇。"
+        f"评述应回答：该文献如何补充机制证据、临床转化线索或竞争格局认知，"
+        f"并指出是否支持 {target} 作为差异化靶点的假设。"
+    )