logic-audit

Compare original and translation side by side

🇺🇸

Original

English

🇨🇳

Translation

Chinese

Logic audit

Logic 审计

Outcome

预期成果

Find and repair material contradictions and missing links across claims, requirements, evidence, actions, and proof. Preserve authority and intent. Don't invent premises, erase real uncertainty, or weaken an exact requirement.

发现并修复在主张、需求、证据、行动和证明之间存在的实质性矛盾与缺失环节。保留权威性与意图。不得凭空创造前提、消除真实不确定性或弱化明确的需求。

Trigger and boundary

触发条件与适用边界

Use when a user asks to find, explain, reconcile, or repair logical inconsistencies, omissions, unsupported claims, conflicting requirements, incomplete cases, or gaps between an objective and its proof.

Use it for prose, policies, plans, specifications, research, decisions, and mixed source sets. For code failures, hand root-cause execution to

systematic-debugging

after this skill identifies the logical contract. For a formal theorem, use a proof assistant or solver when available and limit claims to what it verifies. Routine fact lookup alone doesn't trigger this skill.

A full audit is bounded by the named source set and outcome. Never claim that an open world has no possible gap.

当用户要求查找、解释、调和或修复逻辑不一致、遗漏、无依据主张、冲突需求、不完整案例，或目标与证明之间的差距时使用。

适用于散文、政策、计划、规范、研究、决策以及混合来源数据集。针对代码故障，在本技能识别出逻辑契约后，将根本原因分析工作移交至

systematic-debugging

。对于形式化定理，若有可用的证明助手或求解器则使用它们，并将主张限定在其可验证的范围内。仅常规事实查询不会触发本技能。

完整审计受指定的来源数据集和预期成果限制。绝不能声称开放世界不存在任何可能的漏洞。

Required composition

必备前置组件

Before auditing:

Load
```
always-current-datetime
```
and acquire its clock anchor.
Load
```
starting-point
```
and build its private outcome map.
Load
```
outcome-bounded-work
```
and classify fixed requirements, proof duties, boundaries, forbidden outcomes, mandatory methods, candidate routes, and unknowns.
Keep the same
```
as_of
```
anchor throughout the audit. Reacquire only after a new direct user turn or local midnight.

These skills remain the owners of date handling, the outcome map, and invariant-versus-recipe classification. This skill owns consistency and gap analysis.

审计前需：

加载
```
always-current-datetime
```
并获取其时钟锚点。
加载
```
starting-point
```
并构建其私有成果映射。
加载
```
outcome-bounded-work
```
，并对固定需求、证明职责、边界、禁止成果、强制方法、候选路径以及未知项进行分类。
在整个审计过程中保持相同的
```
as_of
```
锚点。仅在用户发起新的直接交互或本地午夜过后重新获取。

这些技能负责日期处理、成果映射以及不变量与方案分类。本技能负责一致性与漏洞分析。

Evidence setup

证据准备

Name the exact source set, authority order, versions, dates, and excluded material.
Read each source in full when accessible. If truncation remains, mark the omitted range and don't pass the whole source.
Give every material statement a stable source ID in source order.
Classify each statement as definition, fact, assumption, requirement, boundary, exception, evidence, inference, action, conclusion, or unknown.
Normalize terms, entities, references, quantifiers, modal verbs, units, time windows, versions, and scope before calling two statements inconsistent.

明确指定来源数据集、权威层级、版本、日期以及排除的内容。
若可访问则完整读取每个来源。若存在截断，标记省略范围，不得将整个来源视为有效。
为每个实质性语句按来源顺序分配一个稳定的来源ID。
将每个语句分类为定义、事实、假设、需求、边界、例外、证据、推理、行动、结论或未知项。
在判定两个语句不一致之前，先对术语、实体、引用、量词、情态动词、单位、时间窗口、版本和范围进行标准化处理。

Web evidence contract

网络证据契约

Use every available source-relevant web capability with a clear purpose. Do not fire low-level actions that cannot change the finding.

Inventory current search, source extraction, rendered browser, social-source, academic, and domain-source capabilities. Inspect an unfamiliar capability's contract before use.
Keep a capability ledger with one row per distinct source-relevant capability. Mark each
```
USED
```
,
```
NOT_APPLICABLE
```
with a reason, or
```
UNAVAILABLE
```
with the observed failure. Every available relevant capability must be used.
Use the search capability for discovery and query variation. Search official or first-party domains first when a claim has an authoritative owner.
Use the source extraction capability on candidate locations before treating them as evidence. A search snippet is a lead, not proof.
Use the rendered browser capability for JavaScript-only, authenticated, interactive, layout-dependent, console, or network evidence. Follow the host's established computer-use owner.
Use the social-source capability for claims about social posts, profiles, or threads. Use source-specific academic, market, map, repository, or platform capabilities when their corpus can change the finding.
Use a terminal HTTP client only as a recorded fallback after the normal web capability fails or lacks the needed response surface. Never turn a fallback into evidence without inspecting the returned source.
Record URL or source ID, publisher, publication or update date when available, access date, relevant claim IDs, and source limits.

Prefer primary sources. Use independent corroboration when error cost, dispute, or source incentives make one source insufficient. Keep current facts separate from historical evidence.

有明确目的地使用所有可用的与来源相关的网络能力。不得执行无法改变调查结果的低级操作。

盘点当前的搜索、来源提取、浏览器渲染、社交来源、学术来源以及领域来源能力。使用不熟悉的能力前需检查其契约。
维护一份能力台账，每个与来源相关的独特能力占一行。标记每个能力为
```
USED
```
（已使用）、
```
NOT_APPLICABLE
```
（不适用，需注明原因）或
```
UNAVAILABLE
```
（不可用，需注明观测到的故障）。所有可用的相关能力都必须使用。
使用搜索能力进行发现和查询变体。当主张存在权威所有者时，优先搜索官方或第一方域名。
在将候选位置视为证据前，使用来源提取能力处理它们。搜索片段只是线索，而非证据。
针对仅JavaScript、需认证、交互式、依赖布局、控制台或网络证据，使用浏览器渲染能力。遵循主机既定的计算机使用规则。
针对关于社交帖子、个人资料或话题的主张，使用社交来源能力。当特定来源的语料库可能改变调查结果时，使用对应学术、市场、地图、代码库或平台的专属能力。
仅在常规网络能力失败或缺乏所需响应界面时，将终端HTTP客户端作为记录在案的备选方案使用。未检查返回的来源前，绝不能将备选方案的结果视为证据。
记录URL或来源ID、发布者、发布或更新日期（若有）、访问日期、相关主张ID以及来源限制。

优先使用原始来源。当错误成本、争议或来源动机导致单一来源不足时，使用独立佐证。将当前事实与历史证据分开。

Audit procedure

审计流程

Build a dependency map from each conclusion or required outcome back to its premises, actions, owners, and evidence.
Run every applicable consistency class in
```
references/check-catalog.md
```
against each statement, pair, dependency chain, and authority layer.
Run every applicable gap class in that catalog against every required outcome, branch, interface, state transition, and proof claim.
Search for counterexamples and boundary cases before confirming a universal or exhaustive claim.
Verify externally checkable premises through the web evidence contract. Don't use source agreement as a substitute for valid inference.
Create one finding for each independent issue. Don't split one root issue into inflated duplicates.

Adjudicate each candidate as

CONFIRMED

PROBABLE

POSSIBLE

NOT_AN_ISSUE

, or

BLOCKED

Rank impact as
```
critical
```
,
```
high
```
,
```
medium
```
, or
```
low
```
based on the accepted outcome, not on wording intensity.
Repair the smallest coherent unit. Prefer clarification, scope correction, explicit premise, missing case, evidence replacement, or requirement reconciliation over a broad rewrite.
Rebuild the dependency map and rerun the audit on the repaired artifact.

为每个结论或所需成果构建依赖关系图，追溯至其前提、行动、所有者和证据。
针对每个语句、语句对、依赖链和权威层级，运行
```
references/check-catalog.md
```
中所有适用的一致性检查类别。
针对每个所需成果、分支、接口、状态转换和证明主张，运行该目录中所有适用的漏洞检查类别。
在确认全称或穷尽性主张前，搜索反例和边界案例。
通过网络证据契约验证可外部检查的前提。不得用来源一致性替代有效推理。
每个独立问题生成一个调查结果。不得将一个根本问题拆分为多个重复项。
将每个候选问题判定为
```
CONFIRMED
```
（已确认）、
```
PROBABLE
```
（大概率）、
```
POSSIBLE
```
（可能）、
```
NOT_AN_ISSUE
```
（无问题）或
```
BLOCKED
```
（受阻）。
根据已接受的成果而非措辞强度，将影响等级划分为
```
critical
```
（严重）、
```
high
```
（高）、
```
medium
```
（中）或
```
low
```
（低）。
修复最小的连贯单元。优先选择澄清、范围修正、明确前提、补充缺失案例、替换证据或调和需求，而非大范围重写。
重建依赖关系图，并对修复后的工件重新运行审计。

Finding ledger

调查结果台账

Each finding must include:

ID and source locator or statement IDs.
Type and affected outcome or proof duty.
Candidate conflict or missing link.
Normalized reading and competing readings.
Status, impact, and confidence.
Evidence used, including web capability rows.
Minimal repair and any side effects.
Resolution state and remaining uncertainty.

A difference is not a contradiction until both statements have the same relevant entity, time, scope, meaning, authority context, and modality. Missing evidence is not proof that a claim is false.

每个调查结果必须包含：

ID和来源定位符或语句ID。
类型及受影响的成果或证明职责。
候选冲突或缺失环节。
标准化解读及竞争性解读。
状态、影响等级和置信度。
使用的证据，包括网络能力台账条目。
最小修复方案及任何副作用。
解决状态及剩余不确定性。

只有当两个语句的相关实体、时间、范围、含义、权威背景和情态都相同时，差异才构成矛盾。缺失证据并不证明主张为假。

Repair rules

修复规则

Higher-priority instructions and explicit user requirements win.
Preserve the strongest jointly satisfiable contract. Don't silently choose one side of an authority conflict.
When a method is mandatory, use it. Label any alternative as an optional comparison and never substitute it silently.
State new assumptions as assumptions and request approval when they change material action.
Keep unresolved alternatives visible when evidence cannot choose among them.
Separate source correction from implementation. Use the rightful editor and normal side-effect gates.
When rewriting, load
```
meaning-preserving-rewrite
```
and preserve every accepted requirement and proof duty.

优先级更高的指令和明确的用户需求优先。
保留最严格的可共同满足契约。不得在权威冲突中私下选择某一方。
当方法为强制要求时，必须使用。将任何替代方案标记为可选比较项，绝不能私下替换。
新假设需明确标注为假设，且当它们会改变实质性行动时需请求批准。
当证据无法在多个备选方案中做出选择时，需保留未解决的备选方案可见性。
将来源修正与实现分离。使用合法的编辑器和常规副作用管控机制。
重写时，加载
```
meaning-preserving-rewrite
```
并保留所有已接受的需求和证明职责。

Proof threshold

验证阈值

Return

PASS

only when all of these are true:

The full bounded source set was read or every inaccessible range is excluded from the claim.
No unresolved critical or high-impact contradiction remains.
Every required outcome has a complete chain of premises, owner, action or inference, and deciding evidence.
Every material external premise has current source evidence or is marked unresolved.
The web capability ledger accounts for every available source-relevant capability.
The repaired artifact survives the same checks plus at least one counterexample or negative case.
Component checks are not presented as whole-outcome proof.

Use

PARTIAL

when useful findings are proved but the threshold is incomplete. Use

BLOCKED

when missing authority, source access, or evidence prevents a safe repair.

仅当以下所有条件都满足时返回

PASS

：

已完整读取所有受限来源数据集，或所有无法访问的范围已被排除在主张之外。
不存在未解决的严重或高影响等级矛盾。
每个所需成果都有完整的前提、所有者、行动或推理以及决定性证据链。
每个实质性外部前提都有当前来源证据，或已标记为未解决。
网络能力台账已记录所有可用的与来源相关的能力。
修复后的工件通过了相同的检查，且至少通过一个反例或负面案例测试。
组件检查不得作为整体成果的证明。

当已证明有价值的调查结果但未达到阈值时，返回

PARTIAL

。当缺失权威、来源访问权限或证据导致无法安全修复时，返回

BLOCKED

。

Output

输出

Lead with the result and

as_of

date. Then provide the bounded source set, finding ledger, repaired artifact or exact patch, unresolved items, web capability ledger, and final

PASS

PARTIAL

, or

BLOCKED

decision. Cite source IDs and URLs close to the claims they support.

以结果和

as_of

日期开头。随后提供受限来源数据集、调查结果台账、修复后的工件或精确补丁、未解决项、网络能力台账，以及最终的

PASS

、

PARTIAL

或

BLOCKED

判定。在其所支持的主张附近引用来源ID和URL。

Common failures

常见错误

Treating different dates, scopes, definitions, or modalities as contradictions.
Converting ambiguity into a hidden assumption.
Fact-checking premises while ignoring an invalid inference.
Listing gaps without tracing their effect on the outcome.
Using search snippets, source counts, or model agreement as proof.
Calling a repaired paragraph proof that execution works.
Claiming completeness outside the bounded source set.

将不同日期、范围、定义或情态视为矛盾。
将歧义转化为隐藏假设。
验证前提事实却忽略无效推理。
列出漏洞却未追踪其对成果的影响。
将搜索片段、来源数量或模型一致性视为证据。
将修复后的段落视为执行有效的证明。
声称受限来源数据集之外的内容完整。

Package resources

包资源

Use

assets/finding-ledger-template.md

when the user requests a durable ledger. Copy it outside the installed skill before filling it. Read

examples/direct-contradiction.md

examples/temporal-non-contradiction.md

, or

examples/blocked-source-access.md

for complete visible runs. Package maintainers use

references/generation-contract.md

scripts/

scripts/tests/

, and

evals/

, then run

mise run ci

; ordinary audits do not load those paths.

当用户要求生成持久化台账时，使用

assets/finding-ledger-template.md

。填写前需将其复制到已安装技能之外。可阅读

examples/direct-contradiction.md

、

examples/temporal-non-contradiction.md

或

examples/blocked-source-access.md

查看完整的可见运行示例。包维护者使用

references/generation-contract.md

、

scripts/

、

scripts/tests/

和

evals/

，然后运行

mise run ci

；普通审计无需加载这些路径。

Progressive disclosure

渐进式披露

PD-101

references/check-catalog.md

owns the detailed inconsistency and gap taxonomy. Load it before a full audit or when a candidate doesn't fit the core classes. This file owns the trigger, procedure, and backlink.

PD-102

references/eval-cases.json

owns objective pressure cases and acceptance. Load it before testing, reviewing, or changing this skill. This file owns runtime behavior and links back to that evaluation owner.

PD-201

references/generation-contract.md

owns portable package structure, source lineage, examples, and evaluation proof. Package maintainers load it before changing those surfaces. This file owns the public procedure and backlink.

PD-101

references/check-catalog.md

包含详细的不一致性和漏洞分类体系。在完整审计前或候选问题不符合核心类别时加载该文件。本文件负责触发条件、流程和反向链接。

PD-102

references/eval-cases.json

包含客观压力测试案例和验收标准。在测试、评审或修改本技能前加载该文件。本文件负责运行时行为，并链接至对应的评估所有者。

PD-201

references/generation-contract.md

包含可移植包结构、来源谱系、示例和评估证明。包维护者在修改这些内容前加载该文件。本文件负责公开流程和反向链接。