Provenance Weakness: what it measures and what it cannot prove.
Flag numbers, anonymous authority phrases, broad claims, and quotes that need visible sourcing before agents publish, cite, train on, or moderate content.
Plain-English method
Provenance weakness is deterministic claim triage. It does not verify the web; it asks whether the text itself contains enough visible support for claims that would matter in production routing.
Mechanism and scoring
The score adds pressure for unsupported numbers, anonymous authorities, and absolutes, then subtracts visible source cues. This makes it cheap enough for browser-local and offline MCP use.
What this catches
- Unsupported statistical and numerical claims
- Anonymous authority phrases like ‘studies show’
- Overbroad claims that need scope
- Source-cue density relative to claim pressure
What this misses
- Whether a cited source actually supports a claim
- Claims requiring web retrieval or private documents
- Authorship or AI-generation proof
This is one signal in a layered stack.
Single-method detectors are too easy to overtrust. Provenance Weakness is useful when it changes routing: allow, revise, human_review, or reject. It should be layered with specificity, provenance, pattern pressure, Unicode sanitation, media provenance, and paid Deep Scan when the decision matters.
check_provenance_weakness is available through local MCP with no LLM cost, and through the remote MCP endpoint with free unauthenticated rate limits. See all detection methodologies and the dedicated methodology page for the deepest treatment of this signal.