技能 人工智能 构建可审计证据地图

构建可审计证据地图

v20260925
build-evidence-map
该技能用于将有争议的问题或复杂决策转化为结构化、可审计的证据地图。它系统性地追踪所有支持、矛盾、限定和缺失的证据,确保所有推理关系和源证据区域得以完整保留,支持严谨的批判性审查。
获取技能
464 次下载
概览

Build Evidence Map

Turn one contested question into a portable decision artifact that shows what
supports the current position, what pushes against it, and what remains unknown.
Do not use a graph to decorate an answer that has not been sourced.

For a simple factual claim or a general fact-checking request, use a verification
workflow such as doublecheck instead. Use this skill when the relationships
between evidence, intermediate claims, trade-offs, and missing facts matter.

Workflow

  1. Frame one decision. Write one falsifiable question and one provisional

position. Narrow the question until a reader can identify what action or
belief the map is testing.

  1. Collect bounded source regions. Prefer direct observations and primary

sources. Record the URL or absolute local path, publisher, publication date,
retrieval date, section/page/line/timestamp locator, and a short checkable
excerpt. Read references/evidence-ladder.md
when source quality is disputed.

  1. Atomize the reasoning. Create only four node types:
  • position: the single current verdict;
  • claim: an intermediate proposition;
  • evidence: a faithful statement of one source region;
  • unknown: a specific missing fact that could change the verdict.
  1. Type every edge. Use supports, contradicts, qualifies, or

missing. Add a plain-language note explaining why the source node bears on
the target. Topical similarity is not support. Different scope, date, or
population is not automatically a contradiction.

  1. Preserve counterevidence. Do not delete contrary evidence because the

provisional verdict survives it. Represent scope differences with
qualifies edges.

  1. Express uncertainty structurally. Do not invent confidence percentages.

Add an unknown, narrow the position, or qualify a claim.

  1. Write UTF-8 JSON with a .doubt.json suffix. Follow

references/map-schema.md. Keep IDs short,
stable, and semantic.

  1. Validate fail-closed. Resolve

scripts/validate.mjs relative to this SKILL.md, then run it with Node.js
18 or newer:

```bash
node /scripts/validate.mjs decision.doubt.json
```

The bundled validator uses only Node.js built-ins and does not require npm or
network access. Fix every finding before reporting success. Only say the map
is valid when the command exits 0 and prints VALID followed by a
64-character receipt. A file hash, node count, JSON parse, or manual schema
review is not a Doubt receipt. If deterministic validation cannot run, report
that block instead of inventing success.

Render the validated map only when the user has already installed
doubt-ai@0.8.0; do not install or execute a remote package implicitly:

```bash
doubt map decision.doubt.json --out decision.html
```

  1. Verify source snapshots only with explicit network permission. The

following command retrieves each recorded HTTP(S) source and fails closed if
an excerpt cannot be matched:

```bash
doubt verify decision.doubt.json \
--out decision.verified.doubt.json
```

Never run this command implicitly. Local file verification does not use the
network. Do not write a verification object by hand or hide a mismatch.

  1. Inspect the deliverable. Confirm that the question, verdict,

counterevidence, unknowns, edge notes, and exact source regions remain
readable. Treat JSON as the canonical editable artifact; HTML is a
shareable view.

Quality gates

A finished map must satisfy all of these:

  • exactly one position has incoming reasoning;
  • every evidence node names one source and participates in an edge;
  • every source is used and has dates, a bounded locator, and a substantive

excerpt;

  • every non-position node has a directed path to the position;
  • the reasoning graph has no duplicate edges or directed cycles;
  • contrary or qualifying evidence is present when the source set contains it;
  • each decision-changing gap is an explicit unknown node;
  • every edge note explains support, contradiction, qualification, or absence;
  • the verdict is no broader than the evidence.

Deliver the result

Report:

  • the current position in one sentence;
  • the strongest counterevidence or qualification;
  • the most important unresolved unknown;
  • paths to the canonical JSON and any rendered HTML;
  • whether deterministic validation and explicit source verification ran.

Never describe a structurally valid map as proven true. Validation establishes
traceability and graph integrity; source quality and inference quality still
require human review.

信息
分类 人工智能
名称 build-evidence-map
版本 v20260925
大小 10.88KB
更新时间 2026-09-27
语言