Fact Check — Primary Source Claim Verification

Verify factual claims against primary sources using web lookups. Training data recall is NOT evidence.

When to Use

• Backlog items marked UNVERIFIED
• Plugin documentation with uncited claims about tools, APIs, CLI flags
• Skills that reference specific software behavior without citation
• After code review agents flag potential fabrication

When NOT to Use

• Structural findings (broken links, missing files, malformed YAML) — those are filesystem checks, not fact checks
• Code logic bugs — use /dh:root-cause-tracing-process instead
• Research on new tools — use /research-curator instead

Evidence Rules

<evidence_rules>

Adapted from the root-cause-tracing-process evidence chain protocol.

Valid evidence (one of these MUST support each verdict):

1. WebFetch output — content retrieved from an official documentation URL
2. WebSearch result — search results linking to authoritative sources
3. Command output — npx <tool> --help, gh api, or similar CLI output
4. Repository source code — file content from the tool's GitHub repo via gh or clone
5. MCP tool output — mcp__Ref, mcp__exa results with URLs

NOT valid evidence (these MUST NOT support verdicts):

• Training data recall ("I know this from my training")
• Inference from absence ("the docs don't mention it, so it doesn't exist")
• Reasoning from analogy ("tool X works this way, so tool Y does too")
• Documentation that describes intent without observed behavior
• Another AI's claim without primary source backing

</evidence_rules>

Claim Extraction

Parse the input to extract discrete, falsifiable claims:

flowchart TD
    Start([Parse input]) --> Q{Input type?}
    Q -->|Backlog item title| ReadBacklog[Run backlog list --format json, find matching item]
    Q -->|Plugin path| ScanFiles[Scan SKILL.md and references for factual claims]
    Q -->|--all-unverified| FindAll[Run backlog list --format json, filter UNVERIFIED items]
    ReadBacklog --> Extract[Extract falsifiable claims]
    ScanFiles --> Extract
    FindAll --> Extract
    Extract --> Classify[Classify each claim]
    Classify --> Spawn[Spawn verification agents]

Claim Classification

For each claim, determine:

1. Claim text — the specific assertion
2. Primary source — where to check (official docs URL, GitHub repo, CLI help)
3. Verification method — WebFetch URL, run CLI command, search GitHub issues
4. Falsifiability — what would disprove this claim?

Verification Agent Spawning

Spawn @fact-checker agents in parallel waves of 5.

Each agent receives:

CLAIM: {exact claim text}
SOURCE_FILE: {file containing the claim, with line numbers}
PRIMARY_SOURCE: {URL or command to check against}
VERIFICATION_METHOD: {WebFetch | WebSearch | CLI command | gh API}
FALSIFICATION_CRITERIA: {what would disprove this}

Wave Execution

flowchart TD
    Start([Claims extracted]) --> Count{How many claims?}
    Count -->|1-5| Wave1[Wave 1 — all in parallel]
    Count -->|6-10| Split2[Wave 1 — first 5, Wave 2 — remaining]
    Count -->|11+| SplitN[Waves of 5, sequential]
    Wave1 --> Collect[Collect verdicts]
    Split2 --> SeqWaves[Execute waves sequentially]
    SeqWaves --> Collect
    SplitN --> SeqWaves
    Collect --> Report[Generate report]

Verdict Format

Each agent returns a structured verdict:

CLAIM: {the claim being checked}
VERDICT: VERIFIED | REFUTED | INCONCLUSIVE
EVIDENCE:
  - Source: {URL or command}
  - Retrieved: {date}
  - Content: {relevant excerpt from primary source}
EXPLANATION: {how the evidence supports the verdict}
CITATION: {formatted citation for backlog/docs update}

Verdict Criteria

• VERIFIED — primary source confirms the claim. Evidence excerpt matches.
• REFUTED — primary source contradicts the claim. Evidence excerpt shows the discrepancy.
• INCONCLUSIVE — primary source could not be reached, does not address the claim, or is ambiguous. State what additional step would resolve it.

Chain of Verification (CoVe) Requirement

Each agent MUST apply CoVe before finalizing:

1. Initial lookup — fetch primary source, form initial verdict
2. Verification questions — generate 2-3 questions that could falsify the initial verdict
3. Independent check — answer each verification question using a different source or method
4. Final verdict — confirm or revise based on cross-checking

This prevents the agent from confirming its own bias in a single lookup.

Report Format

After all waves complete:

# Fact Check Report

**Date**: {YYYY-MM-DD}
**Scope**: {backlog item title | plugin path | all-unverified}
**Claims checked**: {N}

## Summary

| Verdict | Count |
|---------|-------|
| VERIFIED | {N} |
| REFUTED | {N} |
| INCONCLUSIVE | {N} |

## Verdicts

### Claim 1: {claim text}

**Verdict**: {VERIFIED|REFUTED|INCONCLUSIVE}
**Source**: {primary source URL}
**Evidence**: {excerpt}
**Citation**: {formatted citation}

### Claim 2: ...

## Backlog Updates

{For each REFUTED or VERIFIED claim, the specific edit to make to the per-item file}

Post-Actions

1. Update backlog — for VERIFIED/REFUTED claims, update the backlog item's Status field in the per-item file
2. Lint — uv run prek run --files .claude/backlog/
3. Commit — git add .claude/backlog/ && git commit -m "docs(backlog): fact-check {N} claims ({date})"

For INCONCLUSIVE claims, add a note to the backlog item describing what additional verification is needed.

References

• Evidence chain protocol adapted from root-cause-tracing-process
• Wave spawning pattern from research-curator
• Anti-hallucination checkpoint from skill-research-process
• Chain of Verification from cove-prompt-design