duncanmcqueen wants legal AI answers to come from a committee, not a solo act
This fork turns useful model disagreement into one considered answer, without asking users to run the same prompt repeatedly.
Users could build their own committees in settings, choosing several AI models to address one legal question and a chair model to combine their responses. The committee would appear in the usual model chooser alongside individual options.
- Members would answer in parallel, so waiting time should track the slowest participant rather than every response added together.
- A committee would only be available when the user has access to every member and its chair.
- Deleted or unavailable committees would produce a clear refusal, rather than quietly switching to another model.
- Committee members would not use document or case-law tools, avoiding claims that they completed research they could not perform.
The trade-off is cost: each member is a separate model call, plus the chair's synthesis.
So what Legal teams comparing AI judgement on high-stakes questions should care if they want structured second opinions without losing clarity about what produced the final answer.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?