# 安全治理刷新结论层

本文件把已经完善的证据链上移到顶部结论层。每条结论都列明证据等级、引用资源、公网 URL、决策用途和不得外推边界。

## rc_01_molot_is_direct_mainline

- Level: `L2_direct_public_evidence`
- Claim: MOLOT 已可作为安全专题的直接论文主线，但它证明的是 Python/JavaScript 包 SAST 静态恶意代码检测，不证明 FCD C/C++ 表。
- Evidence refs: `molot_system_card`
- Source URLs: https://arxiv.org/abs/2606.07792
- Why upgraded: MOLOT 论文直接给出 SAST、静态调用图、行为序列、解释映射和 Python/JavaScript 包评估。
- Decision use: 顶部结论应从“候选模型”升级为“直接论文证据主线”。

Blocked extrapolation:
- FCD C/C++ performance table
- LRI Agent Gate runtime/FPR
- independent Open Malicious-Code Bench artifact URL

## rc_02_behavior_chain_becomes_detection_principle

- Level: `L2_multi_paper_supported_principle`
- Claim: 行为链检测已经从方法背景上移为安全专题的核心原则：单个 API/关键词不足以判断恶意目的，必须看 activity chain。
- Evidence refs: `molot_system_card`, `cerebro_behavior_sequence`, `npm_malicious_detection_benchmark`
- Source URLs: https://arxiv.org/abs/2606.07792, https://arxiv.org/abs/2309.02637, https://arxiv.org/abs/2603.27549
- Why upgraded: MOLOT 使用行为序列，CEREBRO 统一 NPM/PyPI 行为序列，NPM benchmark 明确 behavioral chain 能消解单 API 意图歧义。
- Decision use: LRI gate taxonomy 应以 activity_chain 为一等字段，而不是 isolated keyword/API。

Blocked extrapolation:
- binary-only detection
- exact precision without benchmark table extraction

## rc_03_false_positive_budget_is_required

- Level: `L2_direct_operational_constraint`
- Claim: 安全门禁不能只追求召回率；误报预算和 block/ask_human 分流必须前置。
- Evidence refs: `python_malware_benchmark`, `npm_malicious_detection_benchmark`, `molot_system_card`
- Source URLs: https://arxiv.org/abs/2209.13288, https://arxiv.org/abs/2603.27549, https://arxiv.org/abs/2606.07792
- Why upgraded: PyPI benchmark 证明 repository moderation 对误报率极敏感；MOLOT 明确产品约束包含 runtime/memory/FPR；NPM benchmark 支持工具组合和 false-positive tradeoff。
- Decision use: Agent Gate MVP 应默认区分 block、warn、ask_human，而非二元恶意/良性。

Blocked extrapolation:
- claiming LRI measured FPR before logs exist

## rc_04_explainability_is_review_evidence

- Level: `L2_direct_plus_foundation_evidence`
- Claim: SHAP/解释不应停留在可视化，而应上移为 reviewer evidence：source_locations、activity_chain、reviewer_action 必须绑定。
- Evidence refs: `molot_system_card`, `shap_unified_approach`
- Source URLs: https://arxiv.org/abs/2606.07792, https://arxiv.org/abs/1705.07874
- Why upgraded: MOLOT 把可疑行为映射回源码位置；SHAP 提供局部特征归因基础。
- Decision use: reviewer_evidence_schema 应保持为工程化主产物。

Blocked extrapolation:
- SHAP alone as causal proof
- SHAP alone deciding block

## rc_05_graph_pdg_is_secondary_chain_not_molot_metric

- Level: `L1_background_evidence_with_boundary`
- Claim: PDG/Graph 证据可作为源码/片段相似检测的第二检测面，但不能上移为 MOLOT 性能或 FCD 表直接证据。
- Evidence refs: `pdg_holmes_graph_siamese`, `source_code_is_graph`, `fcd_pdg_performance_unresolved`
- Source URLs: https://arxiv.org/abs/2011.11228, https://arxiv.org/abs/2312.16488, user_provided_json
- Why upgraded: PDG/HOLMES 和 CodeGraph 证明图结构在 clone detection 中有公开背景证据；但 FCD 表仍无直接来源。
- Decision use: 工作台应保留 PDG/FCD 线，但所有性能数字必须阻断升级。

Blocked extrapolation:
- verified C/C++ precision/recall/RMSE/FCD speed
- equating graph clone papers with malicious-code detection metrics

## rc_06_agent_gate_is_design_candidate_not_verified_result

- Level: `L0_to_L1_design_candidate`
- Claim: LRI Agent Gate 已具备合同、schema 和 MVP 用例设计，但还不能上移为已验证性能结论。
- Evidence refs: `molot_system_card`, `python_malware_benchmark`
- Source URLs: https://arxiv.org/abs/2606.07792, https://arxiv.org/abs/2209.13288
- Why upgraded: 论文证据足以支撑 gate 设计原则，但缺 LRI 自身运行日志、FPR、延迟和人工复核结果。
- Decision use: 顶部结论应写成“工程设计候选”，下一步是跑 MVP audit log。

Blocked extrapolation:
- verified Agent Gate precision
- verified latency/memory
- verified false-positive budget

## 结论上移日志

- `promoted`: MOLOT as candidate resource → MOLOT as direct evidence mainline，basis=molot_system_card
- `promoted`: behavior sequence as one method detail → activity chain as core gate principle，basis=molot_system_card, cerebro_behavior_sequence, npm_malicious_detection_benchmark
- `partially_promoted`: security gate as generic deployment direction → security gate as contract/MVP design candidate with FPR budget requirement，basis=python_malware_benchmark, molot_system_card
- `downgraded_and_blocked`: PDG/FCD as performance line → PDG/Graph as secondary background; FCD table blocked，basis=pdg_holmes_graph_siamese, source_code_is_graph, fcd_pdg_performance_unresolved
- `blocked_pending_artifact_url`: Open Malicious-Code Bench as available benchmark → paper-stated release only; independent artifact URL required，basis=molot_system_card