---
name: evidence-extraction
description: Extract structured evidence from documents with traceability, quotations, and confidence markers
license: MIT
compatibility: opencode
metadata:
  audience: analysts
  workflow: research
  stage: extraction
---

## What I do

- Read documents and extract evidence relevant to a defined research question
- Separate direct evidence from interpretation
- Capture exact quotations or precise paraphrases
- Record document references for every extracted item
- Structure findings so they can be audited and synthesized later
- Mark uncertain or ambiguous findings for review

## When to use me

Use this after the research task has been scoped and prioritized.

Examples:
- You need requirements, decisions, obligations, controls, or action items extracted from documents
- You need a traceable evidence table
- You need to compare what multiple documents say about the same issue
- You need defensible notes before writing a summary or conclusion

## Inputs I expect

Provide:
- The research objective or sub-question
- One or more documents or excerpts
- Optional triage output from `document-triage`
- Optional extraction schema if the project already has one

## Core extraction rules

For every finding, preserve:

- Finding ID
- Topic or theme
- Evidence type
- Source document
- Section, heading, page, or paragraph reference if available
- Extracted text
- Short analyst note
- Confidence level
- Review flag where needed

Evidence types may include:
- Requirement
- Decision
- Risk
- Assumption
- Responsibility
- Constraint
- Definition
- Date or milestone
- Dependency
- Open issue

## Workflow

1. Read the objective and extraction criteria
2. Scan the document for relevant passages
3. Extract only evidence tied to the research question
4. Distinguish:
   - Direct statement
   - Implied interpretation
   - Missing information
5. Attach traceable references for each item
6. Normalize wording without losing meaning
7. Flag ambiguity, contradiction, or incomplete support

## Output format

Use a structured table like this:

| # | Topic | Evidence Type | Description | Document Reference | Extracted Evidence | Analyst Note | Confidence | Review Needed |
|---|---|----|-----------|-----|--------------------|------|----|---------|------|---------|-------|

Where:
- **Description** is concise and factual
- **Document Reference** is precise enough to locate the source again
- **Extracted Evidence** is a quote or faithful paraphrase
- **Analyst Note** explains relevance without overstating certainty
- **Confidence** is one of: High, Medium, Low
- **Review Needed** is `Yes` when ambiguity remains

## Comparison mode

If multiple documents cover the same topic, add a second table:

| Topic | Source A | Source B | Agreement | Conflict | Notes |
|------|----------|----------|----------|----------|----|

Use this to identify contradictions, differences in wording, or missing alignment.

## Quality bar

- Every claim must be traceable to a source
- Keep extraction atomic: one row per distinct finding
- Preserve important qualifiers such as "must", "should", "may", "unless", and exceptions
- Mark uncertainty explicitly instead of smoothing it over
- Prefer exact citations over memory-based summaries

## Do not

- Do not write the final conclusion here
- Do not collapse multiple findings into one vague row
- Do not omit document references
- Do not present interpretation as if it were a quote
- Do not silently resolve contradictions

## Hand-off to companion skills

After this skill, use:
- `research-synthesis` to turn extracted evidence into a conclusion, report, briefing, matrix, or recommendation