Evidence ChainReview Methodology

Security Governance Evidence Chain

Evidence-backed conclusion chain for Skills/plugin poisoning defense, MOLOT, PDG/FCD boundaries, and LRI Agent Gate design.

Workbench EvidenceMarkdown source

Refreshed conclusions

Promoted conclusions with references and blocked extrapolations.

L2_direct_public_evidence

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。

顶部结论应从“候选模型”升级为“直接论文证据主线”。

https://arxiv.org/abs/2606.07792

L2_multi_paper_supported_principle

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。

LRI gate taxonomy 应以 activity_chain 为一等字段，而不是 isolated keyword/API。

https://arxiv.org/abs/2606.07792, https://arxiv.org/abs/2309.02637, https://arxiv.org/abs/2603.27549

L2_direct_operational_constraint

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。

Agent Gate MVP 应默认区分 block、warn、ask_human，而非二元恶意/良性。

https://arxiv.org/abs/2209.13288, https://arxiv.org/abs/2603.27549, https://arxiv.org/abs/2606.07792

L2_direct_plus_foundation_evidence

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。

reviewer_evidence_schema 应保持为工程化主产物。

https://arxiv.org/abs/2606.07792, https://arxiv.org/abs/1705.07874

L1_background_evidence_with_boundary

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。

工作台应保留 PDG/FCD 线，但所有性能数字必须阻断升级。

https://arxiv.org/abs/2011.11228, https://arxiv.org/abs/2312.16488, user_provided_json

L0_to_L1_design_candidate

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。

顶部结论应写成“工程设计候选”，下一步是跑 MVP audit log。

https://arxiv.org/abs/2606.07792, https://arxiv.org/abs/2209.13288

Evidence quality matrix

Authority, recency, directness and allowed use.

IDAuthorityRecencyDirectnessAllowed use
molot_system_cardpeer_reviewedcurrentdirectcurrent_claim
cerebro_behavior_sequencepeer_reviewedagingbackgroundbackground
python_malware_benchmarkpeer_reviewedagingbackgroundbackground
npm_malicious_detection_benchmarkpeer_reviewedcurrentbackgroundbackground
oss_malicious_package_analysispeer_reviewedcurrentbackgroundbackground
backstabbers_knife_collectionpeer_reviewedagingbackgroundbackground
unveiling_malicious_logicpeer_reviewedcurrentbackgroundbackground
llm_pypi_malicious_detectionpeer_reviewedcurrentbackgroundbackground
pdg_holmes_graph_siamesepeer_reviewedagingbackgroundbackground
source_code_is_graphpeer_reviewedagingbackgroundbackground
shap_unified_approachpeer_reviewedagingbackgroundbackground
fcd_pdg_performance_unresolveduser_inputunverifiedunresolvedcandidate_only