fact-check

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 /find-cause instead
• Research on new tools — use /research-curator instead

Evidence Rules

<evidence_rules>

Adapted from the find-cause 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]