Compare LLM Reasoning
Give several models the same question and context. Compare the assumptions they make, the objections they raise, and how their answers change after review.
A leaderboard cannot answer your question.
Compare responses on the work you actually need to assess.
Benchmarks summarize performance on fixed tasks. They cannot show how a model will respond to your documents, constraints, or review instructions.
Give available models the same case and inspect their responses in one record. The comparison is specific to that review; it is not a general model ranking.
Illustrative example: comparing the assumptions
"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 Challenger 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.