NOMOS 决策审计工作台 · 用户指南NOMOS Decision Audit Workbench · User Guide

1. 产品简介1. Overview

NOMOS 决策审计工作台是一个 确定性、可审计、有责任归属 的决策支持工具。它把一次重大决策拆解为六个可复查的环节：候选方案对比、压力场景、认知审计发现、信息优先级、责任映射与人工审批门，最终输出附着哈希链的审计报告。NOMOS Decision Audit Workbench is a deterministic, auditable, accountable decision-support tool. It decomposes a major decision into six reviewable stages — alternatives comparison, stress scenarios, cognitive audit findings, information priority, responsibility mapping, and a human approval gate — and finally produces a hash-chained audit report.

设计原则Design Principles

• 确定性Deterministic：同一输入必然得到同一输出，结论可复算。: same input, same output, reproducible.
• 可追溯Auditable：每一步评分、约束检查与结论均有记录，并以哈希链固定。: every score, constraint check and conclusion is recorded and fixed by a hash chain.
• 责任闭环Accountable：每个要素都标注负责人与授权来源，避免「集体决策无人担责」。: every element carries an owner and an authorization source.
• 人留最后Human in the loop：审批门之前的一切都是候选，签字才是结论。: everything before the gate is a candidate; the signature is the conclusion.

2. 快速上手2. Quick Start

打开工作台即可开始，无需注册。整个审计闭环分六步：Open the workbench and start immediately. The audit loop has six steps:

3. 核心机制3. Core Mechanics

工作台共 6 个功能卡片，自顶向下构成完整审计流水线。The workbench renders 6 function cards that form a complete audit pipeline.

3.1 候选方案对比3.1 Alternative Comparison

对每个候选方案按多个维度做 加权评分，再叠加 硬约束 与 软约束 检查：Each alternative is scored by weighted dimensions, then checked against hard and soft constraints:

• 维度权重：如示例中规模维度权重 0.45，归一化评分 0.04，加权得分 0.0180。Dimension weights: e.g., scale weight 0.45, normalized score 0.04, weighted score 0.0180.
• 硬约束：不满足即淘汰，例：首个财年预算上限、法务前置条款。Hard constraints: failure means elimination, e.g., first-year budget cap, legal preconditions.
• 软约束：不满足不淘汰，但按惩罚系数扣分（如示例扣 0.10）。Soft constraints: no elimination but a penalty (0.10 in the example).

3.2 压力场景3.2 Stress Scenarios

基于声明场景（如市场准入失败、汇率波动、牌照延期）对候选做 稳健性检验。结构化效应会明确告知受影响候选的资格状态：Under declared scenarios (e.g., failed market access, FX volatility, license delay) candidates undergo robustness checks. The structural effect states eligibility explicitly:

structural_effect → The affected alternatives lose their declared validity
basis; their eligibility must be recomputed before approval.

压力场景命中后，相关替代方案的权重会被重估（示例 adjusted_weight = 0.55），资格需重算后方可进入审批。When a stress hits, affected alternatives get re-weighted (0.55 in the example) and eligibility must be recomputed before approval.

3.3 认知审计发现3.3 Cognitive Audit Findings

认知审计层负责暴露决策过程的结构性弱点，包括三类发现：The cognitive audit layer exposes structural weaknesses in three kinds of findings:

• 最薄弱变量Weakest variable：对总得分影响最大、最不确定的输入。: the input with the largest influence and the highest uncertainty.
• 反证分支Inverse branch：与当前偏好结论相反的证据链，防止确认偏误。: evidence chains that contradict the preferred conclusion, countering confirmation bias.
• 信息黑洞Information gaps：关键变量缺失导致因果链断裂，需补齐条件后才能定论。: missing key variables that break the causal chain; conditions must be filled before concluding.

3.4 信息优先级3.4 Information Priority

决策输入按影响程度分三级排序，帮助团队聚焦关键信息：Inputs are tiered by impact to focus the team:

3.5 责任映射3.5 Responsibility Map

责任映射将每个决策要素绑定到「负责人 + 授权来源」，二者缺一即视为责任缺口：The responsibility map binds every element to an owner and an authorization source; a missing either side is a gap:

这一设计将责任追溯到最小决策单元，杜绝「人人都有责、人人无责任」。Accountability traces back to the smallest decision unit — no diffusion of responsibility.

3.6 审计报告与哈希链3.6 Audit Report & Hash Chain

每一步审计事件都会写入哈希链：每个事件哈希由前一事件哈希参与计算，任何单点改动都会让整条链失效。核心结构：Every audit event enters a hash chain: each event hash incorporates the previous event hash, so any single alteration invalidates the whole chain. Core structure:

algorithm_audit_root_hash = e890018e8d7a261b948b03f02f03d2f2f48cddfc71b3d67c8108f34c71825d99
event_hash        = b67942293097de189e89d7cb8023d5fe99ba3964770e39b0b151072186b18f01
previous_event_hash = b67942293097de189e89d7cb8023d5fe99ba3964770e39b0b151072186b18f01
record_hash       = 8b0e373aebdcf77efc05c7f495868155349c33bd320965753fb5912e130a7dae
report_hash       = d192f11aaede5426248ae7e819d5f4e2d9469d4dc724fc41ca369c526b546299
reconstruction_id = REC-499454CC7F74

审计报告以自然语言给出审查结论，并携带哈希值供第三方复核——报告可以被任何人独立验证，但无法被伪造。The report gives a narrative conclusion with hashes for independent verification — anyone can verify, nobody can forge.

3.7 人工审批门3.7 Human Approval Gate

流水线终点是人工审批门。只有授权人签字后，报告才正式生效；若被驳回，系统按反事实逻辑注入修正变量并重跑候选集，收敛后再次提交人工裁决。The pipeline ends at the Human Approval Gate. The report takes effect only after an authorized signature; if rejected, the system injects corrective variables, reruns the candidate set, and resubmits for a new human ruling.

• 签字生效：结论由人承担，引擎不背锅、也不揽责。Signature finalizes: humans own the outcome.
• 驳回重跑：修正变量收敛后等待再次裁决。Rejection reruns: corrected variables converge, then a new ruling.

4. 常见问题4. FAQ

Q1：数据会上传到服务器吗？Q1: Is my data uploaded?

演示场景中所有候选、评分与审计事件均在本机浏览器内运行，不上传任何服务器。In the demo, all candidates, scores and audit events run locally in your browser — nothing is uploaded.

Q2：哈希链能做什么用？Q2: What is the hash chain for?

用于事后核验审计过程未被篡改：重新计算事件哈希并与链上记录比对即可；也可作为仲裁或举证材料。To prove the audit process is untampered: recompute event hashes and compare — usable as arbitration evidence.

Q3：如何更换决策目标？Q3: Can I change the decision goal?

修改目标后建议重置候选与评分，再重跑压力场景与认知审计，确保结论基于一致前提。After changing the goal, reset candidates and scores, then rerun stress and cognitive audits for a consistent premise.

Q4：谁负责审批签字？Q4: Who signs off?

由本页面责任映射中标注的负责人执行；系统不代签、不默认通过。The owner recorded in the responsibility map signs; the system never signs on anyone's behalf.

5. 术语表5. Glossary