Skip to content

Latest commit

 

History

History
46 lines (32 loc) · 2.02 KB

File metadata and controls

46 lines (32 loc) · 2.02 KB

Agent-Cite

Agent-Cite is a Claude Code plugin that enforces evidence protocol on AI-generated output. Every factual claim must cite a source, every number must have a derivation, zero uncited inference.

Architecture

Three command journeys, each with a skill and supporting agent:

Command Skill Agent Purpose
/cite-audit evidence-audit citation-scanner Detect uncited claims
/cite-fix citation-fix citation-fixer Fix violations with real citations
/cite-report evidence-report source-verifier Project-wide evidence health

Violation Types

Type Default Severity Meaning
UNCITED_INFERENCE error Factual claim without citation
UNVERIFIED_NUMBER warning Number without derivation
UNSUPPORTED_ABSENCE warning Absence claim without search evidence
BROKEN_CITATION error Citation to non-existent source
FALSE_ABSENCE critical Agent claimed something doesn't exist, but it does
UNVERIFIABLE human decision No source found anywhere — claim may be fabricated

Verdicts

  • COMPLIANT: Zero error-level violations
  • NON_COMPLIANT: Any error-level violation

When to Apply

Apply to AI-generated reports, analyses, documentation, recommendations — any output that asserts facts. Do NOT apply to pure code, formatting, or creative output.

Configuration

evidence-protocol.yaml in project root or .cite/ directory:

  • Custom violation severities
  • Citation pattern formats
  • Include/exclude file patterns
  • Source directories for /cite-fix
  • Health thresholds for /cite-report

Relationship to Agent-PROVE

Agent-PROVE includes evidence auditing as one of its capabilities (/audit command). Agent-Cite is the standalone version focused exclusively on evidence enforcement with additional capabilities (citation fix, project-wide reporting, configurable protocol). If you use Agent-PROVE, you already have basic evidence auditing — Agent-Cite goes deeper.