Claim Grounding Auditor

Audits any AI draft for unsupported claims — flags each one, grades its source, and returns a substantiation report.

Install
cmdop skills install agensi-claim-grounding-auditor

Claim Grounding Auditor is a skill that takes a draft and its source material, extracts every checkable factual claim, and returns a structured substantiation report. Each claim receives exactly one of four statuses: GROUNDED (directly supported by a strong, citable source), INFERRED (a conclusion assembled across sources rather than stated by any one of them), WEAK (supported only by an outdated, low-quality, or non-authoritative source), or UNSUPPORTED (no source found, or the source contradicts the claim).

The skill enforces a source-strength hierarchy. Primary, authoritative, and current sources rank highest. Aggregated, tertiary, or marketing content ranks lowest. A claim backed only by weak sources cannot be elevated above WEAK regardless of how confident the prose reads.

A core distinction the audit enforces is fact versus inference. When a draft combines two sourced facts into a new conclusion, that conclusion is classified as INFERRED rather than GROUNDED — even when both inputs are solid. This surfaces the most common way confident AI output smuggles in unverified conclusions.

The skill is domain-agnostic and applies to research briefs, RAG output, product copy, technical documentation, analyst reports, and any text where a wrong-but-confident sentence carries real cost. It does not fetch the open web on its own; it audits against sources the agent provides. It checks grounding and source strength, not domain correctness — a claim well-supported by a flawed source passes as GROUNDED against that source.

Use cases

  • Gate RAG output before delivery — flag any claim not grounded in the retrieved documents
  • Audit analyst or research briefs for unsupported or inferred claims before distribution
  • Pre-flight marketing and product copy for statements that legal or compliance teams would challenge
  • Check technical documentation against a codebase or spec for unsupported assertions
  • Run as a final quality gate in a multi-agent writing or summarization pipeline
  • Identify smuggled inferences in LLM-generated summaries that combine facts into unverified conclusions

When to use it

  • The agent produces output that other people act on — reports, docs, summaries, copy
  • A RAG pipeline needs a grounding check before results reach end users
  • A writing workflow requires an auditable, claim-level paper trail
  • The cost of a single fabricated number or unsupported statistic is significant

When not to use it

  • No source material is provided — without supplied sources the skill cannot verify external facts
  • Domain correctness is the goal — the skill grades grounding against sources, not whether the sources themselves are accurate
  • Web retrieval is required — the skill does not fetch the open web unless the agent separately provides that capability
  • The task is creative or fictional writing where factual grounding is not relevant