Find claims that still need evidence.
Ask other models to identify unsupported claims, contradictions, and missing context. Use the review to decide which statements to check against primary sources.
A second model is still not a source.
Treat model feedback as a lead for verification.
An answer can mix supported facts, reasonable inferences, and speculation without marking the difference. Asking the same model for reassurance may repeat the same framing.
Other reviewers can question individual claims and suggest alternative readings. Their disagreement tells you where to investigate; agreement does not verify a statement.
Illustrative example: catching a missing assumption
"The car wash is 100m away. Should I walk or drive there?"
"You should walk to save emissions and get some exercise. It's only 100m!"
"I strongly disagree. This treats the question as a pure transportation optimization problem. If you walk, the car sits unwashed. The correct answer is unequivocally drive."
"The Detector correctly identified the premise's trick. The user needs the car at the destination to utilize the service. A car wash cannot wash a pedestrian."
"Consensus reached. While the initial response prioritized safety and emission conditioning, it failed basic logical constraints. You must drive to get the car washed."
Built for consequential work
Surface unsupported claims
Independent model responses can surface contradictions, missing support, and claims that need external checking before you rely on them.
Use independent perspectives
Assign different roles and review paths so a draft faces more than one perspective. Agreement remains a signal to inspect, not a guarantee of truth.
Inspectable review records
Keep the transcript of the review, including its objections and conclusion, so you can inspect the reasoning and decide what still needs human verification.
How a review works
Bring the work, choose the kind of feedback you need, and inspect the result.
Add the work
Paste a proposal, plan, claim, or AI answer. Include the question and context the reviewers need.
Choose the review
Select models, a review format, and the number of rounds. Follow their responses as the review runs.
Read and revise
Read the conclusion, unresolved objections, and transcript. Revise the work and check the evidence that matters.
Published research
Agreement is useful evidence, not proof of accuracy.
Our observational disagreement study examines 558 completed reviews and 9,507 assessed claims in the product corpus. It describes how model positions changed; it is not an accuracy benchmark.
Read the study →Read public reviews
Open a shared record to inspect the question, model responses, and conclusion.
Public review record
Walk or drive to the car wash?
Inspect one completed review, including the original question and the responses behind its conclusion.