The project

AI models that
check each
other's work.

Mixture of Models explores how AI models can check each other's work. It runs manually today, using prompts and saved model rounds. The planned ring gives models turns on a shared draft and records claims, disagreements, and checks.

Explore the project See the planned ring

Independent perspectives.A shared draft.

01 / Where the project stands

A method today. A system next.

Updated October 8, 2026

Working now

The manual method

Prompts and saved model rounds, with a person moving the work between models and deciding what to keep.

Planned architecture

The review ring

Models take turns on a shared draft, keeping a record of individual claims, disagreements, and checks.

Next milestone

A small runnable prototype

Connect the turn-taking loop and claim record to one checker tool, then inspect a complete run.

02 / The purpose

A second perspective needs a purpose.

A useful review should point to something specific: a missed requirement, an unsupported claim, or a disagreement that needs a check. The project is about organizing that work and keeping the reasons behind the result visible.

01 / The work

What needs checking?

Start with the task, its requirements, and the claims that affect the result.

02 / The review

Where do models disagree?

Keep independent starting answers and record the issues raised during review.

03 / The decision

What holds up?

Use sources and direct checks where applicable, with human judgment for the final decision.

03 / The planned ring

One draft.
Many perspectives.

First, models answer independently. Then they take turns reviewing a shared draft and its claim record. Applicable tool checks and unresolved questions stay with the work.

Explore the ring

Planned architecture. Models take turns; checks and open questions stay attached.

04 / From the Notebook