[PR-1] M1 显著性数值修复 + EvidenceLedger 最小结构

[justDance-everybody]

本任务属于 Epic #23。共享上下文(整体目标、EvidenceLedger 完整 schema、Agent Execution Contract、Non-goals、回归基线)见 #23,不在此重复。
依赖:无(第一个 PR)。
代码定位按当前仓库版本核对;若行号漂移,以函数名 / 字段名 / 挂载点语义为准,不得因行号变化改变任务目标。

完成本 PR 需遵守 Epic #23 的 Agent Execution Contract 与 Test Execution Protocol:先写失败测试与反例夹具 → 最小实现 → 跑里程碑测试 → 跑回归基线(pytest tests/test_pipeline_contracts.py、pytest tests/test_vnext_manuscript.py)。PR 描述需列出:changed files / tests added / tests run / before-after / non-goals。

---

目标

两件事:① 修掉显著性的子串误判;② 建立 EvidenceLedger 的最小结构 + builder(不要求 full traceability,那是 PR-2)。

Allowed changes

• agents/paper_completeness.py
• significance alpha 配置读取相关文件
• EvidenceLedger builder 的最小对象(build_evidence_ledger,见下)
• PR-1 fixtures
• PR-1 tests

Forbidden changes

• 不修改 contracts/pipeline.py
• 不改变 :699 的存在性检查语义
• 不通过修改测试来绕过 p=1.0000 反例
• 不新增从 manifest.statistical_tests 文本解析显著性的路径
• 不在本 PR 实现 assert_traceable / full traceability(留 PR-2)

M1 现状(bug,已定位三处)

• paper_completeness.py:241-242:summary["statistical_tests"] = f"p={packet['p_value']}" —— 把数值塞进字符串。
• paper_completeness.py:589,591:p_text = _lower(manifest.get("statistical_tests")) → has_significance = "p=" in p_text or "bootstrap" in p_text or "permutation" in p_text。p=1.0000 含 "p=" 子串即判显著。
• :699(存在性检查)、:753(显著性问答)也读 statistical_tests 文本。

M1 目标行为

• has_significance 改为读数值:p_value is not None and p_value < alpha,数值来源 ledger 的 p_value / alpha。
• verdict 绑定:verdict ∈ {confirmed, reproduced} 才允许显著性支撑 Abstract 级断言;verdict ∈ {refuted, inconclusive} 一律不显著、不得进 Abstract。
• α 默认 0.05,新增 env DEEPGRAPH_SIGNIFICANCE_ALPHA 可覆盖(沿用现有 DEEPGRAPH_* 风格)。
• 699 与 753 分别处理: :753 是显著性判定,改为按数值判;:699 是"是否记录过检验"的存在性检查,语义不同,本 PR 不动(改成 p<α 会把"记录了但不显著"误报成"缺失")。

取数路径

build_claim_evidence_matrix 内 packet 已在 scope(paper_completeness.py:586 经 _packet_parts(state)),p_value 就在 packet(packet.get("p_value"))。直接读即可,无需改签名。PR 描述说明数值取自 packet。

EvidenceLedger 最小 builder(本 PR 范围)

按 Epic #23 的 schema,实现:

def build_evidence_ledger(packet: dict, benchmark_summary: dict, *, alpha: float,
                          provenance: dict | None = None) -> dict:
    """Build the single source of truth for presentation evidence."""

最小子集需含:alpha / verdict / p_value / effect_size / confidence / per_method / seed_variance / seeds(字段语义与来源见 #23 schema)。validate_evidence_ledger / assert_traceable 留 PR-2。

验收用例(单元,纯 CPU)

#	输入	期望输出
M1-1	p_value=1.0000, verdict=inconclusive	has_significance=False;"Improves utility" 行 can_appear_in_abstract=False
M1-2	p_value=0.0123, verdict=confirmed,manifest.seeds ≥3 项、per_method ≥2 键	has_significance=True;该行 can_appear_in_abstract=True(591 行还要 len(manifest["seeds"])>=3,夹具须同时满足)
M1-3	p_value=0.0123, verdict=refuted	can_appear_in_abstract=False(verdict 否定)
M1-4	p_value=0.049,DEEPGRAPH_SIGNIFICANCE_ALPHA=0.01	has_significance=False(0.049 ≥ 0.01)
M1-5	p_value=None	has_significance=False,不报错
M1-6(回归)	manifest.statistical_tests="p=1.0000" 文本仍在,但数值 p_value=1.0	以数值为准判不显著(证明不再走子串)
M1-7	753 显著性问答:p_value=1.0	答 "No";699 存在性语义保持不变

反例夹具: fixtures/m1_p_eq_one.json、fixtures/m1_significant.json(含 seeds:[0,1,2]、per_method 二键)、fixtures/m1_refuted.json。

DoD

M1-1~7 全绿;591 与 753 改数值;DEEPGRAPH_SIGNIFICANCE_ALPHA 落地;最小 build_evidence_ledger 可用;回归基线全绿;收敛分支、未确认不强推。

[Protocol-zero-0]

Completed in PR #29 on branch epic-23-evidence-ledger.

Changed files:

• agents/paper_completeness.py
• tests/test_paper_completeness_m1.py
• tests/fixtures/m1_p_eq_one.json
• tests/fixtures/m1_significant.json
• tests/fixtures/m1_refuted.json

Tests added:

• M1-1 through M1-7 coverage for p=1.0000, confirmed significant p-value, refuted verdict, DEEPGRAPH_SIGNIFICANCE_ALPHA, missing p-value, text p=1.0000 regression, reviewer significance answer, and minimal build_evidence_ledger.

Commands actually run:

• Pre-change baseline: pytest tests/test_pipeline_contracts.py -> 12 passed
• Pre-change baseline: pytest tests/test_vnext_manuscript.py -> 5 passed
• Milestone: pytest tests/test_paper_completeness_m1.py -> 7 passed
• Regression: pytest tests/test_pipeline_contracts.py -> 12 passed
• Regression: pytest tests/test_vnext_manuscript.py -> 5 passed

Non-goals respected:

• Did not modify contracts/pipeline.py.
• Did not change the statistical-test existence gate semantics.
• Did not implement assert_traceable in this commit; that is in [PR-2] M4 EvidenceLedger full traceability checker #27.

[Protocol-zero-0]

Done via #29 (squash-merged into master 2026-06-06). 验收用例与回归基线均在 PR #29 全绿。