Fugue

Planned

One box.Several voices.

Planned: you type a task once, and you do not pick a model, draw a workflow, or watch a group chat.

How it works

How it works

A fugue is several voices and one piece. The design uses that shape. In the test, models from different providers work as separate voices. The main view would show one deliverable.

The design keeps model choice, workflow diagrams, and a chat between models off the surface.

In the version under test, a second model checks the draft against the material, sentence by sentence. That material is the documents, tests, or data for the task. The check compares each sentence with those sources. It does not offer a competing opinion.

In the test, sentences the sources do not support are removed or flagged. The design returns one deliverable, with its sources. The design keeps disagreement and the working process in voice tracks that stay closed until you open them.

The design aims for a result the reader can verify.

Illustration of the planned box. It is not a live control.

Planned shape of the output. Not a sample result.

Deliverable

  • One memo
  • Sources where a sentence has them
  • A note of what was removed, flagged, or not checked

Voices

Planned: each voice works apart. The voices would not share one context. They would pass pieces of work across, and the main view would still hold one deliverable.

Planned: players can be swapped. This test uses model families such as Grok, Gemini, and Kimi. Those names are examples of players, not a partnership or an endorsement.

Planned: a simple question can stay with one voice, and several voices are for work that has sources, tests, or data outside the models.

Planned: voice tracks stay closed by default, and you open one on demand.

Draft voice

In the test, the first voice writes from the task.

Check voice

In the test, a second voice reads the sources sentence by sentence and marks each sentence supported, contradicted, or absent from the material.

Change record

Planned: the track keeps what was removed or flagged, and what the check did not cover. Opening it would not add a second answer to the main view.

The check

In the test, a supported sentence is one the sources contain. A contradicted sentence is one the sources oppose. An absent sentence is one the sources do not contain.

In the test, removal and flagging follow those marks, and a rule outside the models applies them. The reader is not asked to pick a winning voice.

Planned: the deliverable names the material in scope and marks what was not checked.

Cross-checking between models is not proven to remove errors. Fugue does not claim that it does. No test result is published yet.

Status

Planned: Fugue runs on your own machine, writes artifacts to disk, and lets players be swapped.

Planned: a remote player sends its request to that provider. The workspace, and the artifacts it saves, would stay on the machine.

The first use under test is a fact-check memo: claims checked against a frozen set of public material, returned as one memo with its sources.

This is an MVP. It is still being tested. Nothing has been proven yet.