TECHNOLOGY

Disagreement is the measurement.

Most AI tooling treats model disagreement as an error to suppress. We treat it as the only cheap signal available about whether an answer can be trusted.

A single model is fluent when it is right and fluent when it is wrong, so its confidence tells you very little. Asking that same model to check itself returns the same blind spots with more reassurance attached. Two models from rival vendors do not share a training pipeline, and where they diverge is where you should look.

The verification methodA question enters. Two rival-vendor models answer independently and attack each other's reasoning. A judge rules, producing a verdict, a confidence score and a disagreement map.InputA questionExpensive to get wrongSide AFrontier model, vendor oneAnswers independentlySide BFrontier model, rival vendorAnswers independentlyattack roundsJudgeImpartialRules on what survivedVerdictConfidence scoreDisagreement mapThe dissent is published alongside the ruling. Where the two sides split is the part worth reading.
The verification method: independent answers, structured attack, a judge, and a published dissent.

A ruling, not an average

Averaging two answers hides the disagreement, which is the part worth having. A judge produces a ruling and the dissent stays attached to it.

A confidence score that means something

The score reflects how much survived contest, not how fluent the output sounded. It is a probability raiser with receipts, not an oracle, and we say so on the product site too.

A disagreement map you can hand to someone

The output is designed to be forwarded. Where the models split, what each claimed, and how it was resolved, in a form a colleague or a regulator can read without taking your word for it.

Verification and memory: two layers, one positionA verification layer makes AI outputs checkable and a memory layer makes AI systems continuous. Both sit between the systems people use and the models underneath, with room reserved for further products.WHERE THE WORK HAPPENSAssistants, editors, and the applications an organisation already runsPRODUCT 01 · LIVEVerification layerRival-vendor models argue a claim. A judge rules.Output: a verdict, a confidence score, a disagreement map.Makes an output checkable.PRODUCT 02 · IN DEVELOPMENTMemory layerHosted memory across many AI clients.Privacy and non-mixing are the design constraints.Makes a system continuous.RESERVEDThe architecture holds further products without a redesign.Underneath: frontier models from rival vendors, reached through official APIs. We are not a model vendor and we are not owned by one.
Verification and memory: two layers, one position. Outputs you can check, systems that stay continuous.

What we do not publish.

We publish the method and the output format. We do not publish model routing, prompt strategy, model versions or unit costs. Model versions in particular change on vendor timelines, and a page that names them is wrong within a quarter. This is a policy, not an omission.

Questions about the method?

Contact