NOMOS Decision Engine
确定性决策引擎(NOMOS 0.7.0):十算子认知审计、二阶因果交互、形式化收敛、企业控制平面与哈希链审计。企业级部署需独立服务与数据库。
NOMOS Intelligent Decision Engine
确定性因果推理与可审计决策编排层,不含概率黑箱。每一次决策都可分解为可追溯的因果拓扑,支持合规级审计轨迹、场景压力测试与人在环治理。当前引擎版本 0.7.0(src/second_perspective/version.py)。
Trigger Scenarios
Trigger only when the user explicitly mentions the following named entities (avoid false triggers from broad keywords):
- NOMOS Decision Engine / NOMOS Decision-Hub / second_perspective
- References to the "Sealed Report" specification or "Decision Scenario Library" templates generated by this skill
- Explicit requests for causal counterfactual analysis, root cause tracing, scenario stress testing that specify using this engine
Note: General topics (e.g., "AI decision compliance audit", "enterprise deployment solutions", "high-risk decision governance") should not automatically trigger this skill unless the user explicitly references the above named entities.
版本能力矩阵
| 版本 | 新增能力 |
|---|---|
| 0.7.0 | 元因果账本术语对齐(破坏性):meta_ledger.bases 五键改为 trace_gap / parallel_convergence / narrative_occlusion / bounded_audit / causal_succession;reason 码前缀收敛为 TRACE_ / CONVERGENCE_ / NARRATIVE_ / AUDIT_ / SUCCESSION_;不保留别名 |
| 0.6.0 | 认知审计内核 5 → 10 算子;拓扑底座(四进程校验,含公理 5 链内时序);META 元因果账本(只度量、恒 T2、结构上不可阻断);自主进化层 evolve 只提案不适用(auto_applied ≡ 0);产物版本串 GCAE-1.0.0 → GCAE-1.1.0 |
| 0.5.0 | 企业控制平面:持久化事件存储、租户隔离、OIDC + 作用域强制、KMS 签名、限流、可观测性、迁移/备份、领域控制包。全部 env 门控、默认关闭 |
| 0.4.0 | 二阶因果:假设交互层(I-1~I-4)、形式化收敛(五态)、LLM 三层权限闸门(T1/T2/T3) |
| 0.3.0 | IntelligentDecisionHub 编排器、密封 HubReport、POST /v1/hub/analyze、场景压力测试、信息优先级队列 |
完整变更日志见 CHANGELOG.md;v0.4 架构见 docs/INTELLIGENT_DECISION_HUB_V0_4.md。
十算子认知审计内核
src/second_perspective/audit/cognitive/ 为纯 Python、零依赖、确定性管线(默认无 RNG、无 LLM 调用):
| 算子 | 名称 | 作用 |
|---|---|---|
| ⊙ ORI | Origin Anchor | 原点事件 / 目标状态 / 能量约束 |
| ⊗ NS | Narrative Strip | 剥离修辞、情绪与道德外衣 |
| ⊕ IAP | Implicit Assumption | 显式化未声明前提 |
| ⊿ LCH | Latch / Fragility | 定位最弱变量,计算 Delta D |
| ⊞ TPG | Rule-Free Thinking | 构建 □/→/⦿ 拓扑 + 四路并行校验 |
| BFC | Binary Fact Check | 命题归约为真/假,不引入第三值 |
| ⚙ CCS | Causal-Chain Sync | 反事实与反向校验 |
| ⇄ GRF | Gray Feedback | 灰度阶梯与现实验伪对齐 |
| META | Meta-Causal Ledger | 混沌 / 无极 / 幻象(A6) / 天道(A10) / 轮回;只度量,永不阻断 |
| ⊚ STATE | State Anchor | 锚定责任与最终裁决 |
拓扑底座(.plugins.topology)提供 □/→/⦿ 语法与四路并行校验:一致性 · 约束 · 闭合 · 时序。
ORI 仅在声明原点时注册,GRF 仅在声明灰度阶梯时注册——条件注册,杜绝永恒真阻断。
Core Capabilities
- 确定性评估:硬约束阈值 + 显式软约束惩罚分,无隐藏调分
- 细粒度算法审计:每次主要操作生成哈希链审计事件(
sealed_hash),任何改动都会破坏校验 - 因果反事实重选:假设失效时计算传递失效闭包,重选存活候选
- 二阶交互:协同 / 冗余 / 退化三类交互,四条不变式(非臆测、阶数守恒、效应有界、单调性)
- 形式化收敛:
FIXED_POINT/NO_GAIN/BUDGET_EXHAUSTED/DIVERGED/BLOCKED;BUDGET_EXHAUSTED不是收敛 - 场景压力测试:指标覆盖、证据缺失、假设失效
- 反向根因追溯:从观测偏差反推根因假设,输出因果链
- 敏感性 / 稳健性:Pareto 前沿、脆弱准则、排序稳定性
- LLM 护栏:T1 标注 / T2 提案 / T3 叙述;LLM 输出永不成为裁决
- 企业控制平面:事件存储、租户隔离、OIDC、KMS 签名、限流、可观测性(env 门控)
Usage
安装
cd nomos-decision-hub # 本技能目录(含 pyproject.toml 与 src/)
python -m venv .venv
source .venv/bin/activate # Windows: .venv\Scripts\activate
pip install -e .
命令行
nomos-demo # 核心决策内核摘要
nomos-demo --decision examples/market_entry.json --json
nomos-hub-demo # 智能决策中枢(含两个压力场景)
Basic Decision Analysis
from second_perspective import IntelligentDecisionHub
from second_perspective.models import HubAnalysisRequest
request = HubAnalysisRequest.model_validate({
"decision": decision_payload, # See examples/market_entry.json
"scenarios": [
{"id": "SC1", "name": "Critical assumption failure",
"failed_assumption_ids": ["A1"]},
{"id": "SC2", "name": "Cost shock",
"metric_overrides": {"S2": {"capital_required": 6000000}}},
],
})
report = IntelligentDecisionHub().analyze(request)
print(report.model_dump_json(indent=2))
HTTP API
uvicorn second_perspective.api.main:app --port 8000
# 端点:POST /v1/hub/analyze;兼容保留 v0.2 端点;GET /health;GET /v1/metrics
Enterprise Deployment
# Docker deployment
docker build -t nomos-hub .
docker run -p 8000:8000 \
-e SP_ENV=production \
-e SP_API_KEY=your-secret-key \
-e SP_DATABASE_DSN=postgresql://user:pass@db:5432/nomos \
nomos-hub
生产环境 API key 缺失时 fail-closed 拒绝启动。
技能层扩展(P0-P2)
🔏 Sealed Report
决策输出可封装为防篡改密封报告:哈希链(Merkle 因果链)+ 确定性 JSON + 可选 HMAC 签名。审计方可一键验证(SEAL VALID / TAMPERED)。规范见 references/SealedReportSpec.md。
引擎原生对应物:每个审计事件带 sealed_hash,链根可独立复核;v0.5 起支持 KMS 分离式签名(X-Nomos-Kms-Signature)。
📚 Decision Scenario Library
4 类预置场景模板(合规准入 / 战略押注 / 根因追溯 / 资源配置),复制-替换-运行即得密封报告。见 references/DecisionScenarioLibrary.md。
🌍 Multilingual Export
报告带 lang 字段(zh/en),同一因果链无需重跑引擎即可渲染为目标语言。
Files
| 路径 | 内容 |
|---|---|
SKILL.md | 本文件 |
README.md / README-zh.md | 英文 / 中文说明 |
LICENSE | 分级授权协议(个人研究免费;商用需书面授权) |
CHANGELOG.md | 0.1–0.7 变更日志 |
GLOSSARY.md | 中英术语表 |
src/second_perspective/ | 引擎源码(可 pip install -e .) |
docs/ | v0.2 基础、v0.4 二阶因果、企业模块、验证指南、四项合规文档;docs/archive/ 为 v0.3 历史文档 |
examples/ | market_entry.json、tc1_non_speculation.json、tc3_hard_constraint.json、audit_spec_adoption.json |
openapi-action.yaml | OpenAPI 契约(0.7.0) |
scripts/ | export_openapi.py、migrate.py、backup.py |
references/SealedReportSpec.md | 密封报告格式与验证流程 |
references/DecisionScenarioLibrary.md | 决策场景库 |
Dockerfile / docker-compose.yml | 容器化部署 |
Determinism Contract (Invariants)
- The engine never guesses missing weights, evidence, thresholds, or authorization relationships
- Hard constraints determine eligibility, soft constraints must have explicit penalty scores declared, no silent score changes
- All policies affecting behavior carry
policy_idandversion, embedded in results - Evidence quality is evaluated by designated responsible nodes, the engine does not fabricate credibility
- Assumption failures propagate along the dependency graph, explicitly naming affected alternatives
- Output is always "leading candidate under declared inputs", final decision authority lies outside the algorithm
- Approver name +
authorization_refmust match the anchored decision owner - All evaluations/approvals are new revisions linked via hashes, no silent overwrites
BUDGET_EXHAUSTEDmust never be reported as "converged"- LLM output may only occupy T1 (annotation) / T2 (proposal) / T3 (narrative); it can never flip a status, adjudicate, or alter weights
- META is a measurement tier: it can never block a derivation on its own
Licensing
本作品采用分级授权(见 LICENSE,v1.2,2026-10-14):
- 个人研究许可:自然人可免费用于非商业学术研究、学习或个人实验。
- 商业 / 政府使用:需取得 Shanghai Linming Junhua Technology Co., Ltd. 与 NOHN AI TECHNOLOGY PTE. LTD. 的书面授权。
本作品不是 MIT-0,也不属于公有领域。 若发布平台的授权字段与此不符,以本仓库 LICENSE 为准。
Related skills
- Downloads
- 1.1kAll time
- Security audit
- Review
- Last updated
- 1mo ago
- Current version
- v2.1.0
- License
- MIT-0
