The Debater
Three AIs argue over your idea, vote on every step, and hand you only the plan they agree on.
Stage: closed alpha opens 4 December 2026, by invitation from the waitlist. Plans and progress public; code private.
The question
Can three AIs, working as a small team with fixed rules, turn a rough idea into a plan that's better than any one of them would write alone, and can every step of how they got there be shown?
Concept
You give a topic. Each AI breaks it into a tree of small, testable tasks. The AIs review each other's tasks with a yes, maybe or no vote, and a fixed tally decides what survives. The process judges nothing itself: the AIs judge, the rules count.
Design
- A tree per proposal. One concept at the root, branches for the parts, and leaves for the tasks. Only the leaves are voted on.
- An atomic test for every task. One sentence with no "and"; one owner and one checkable outcome; a pass-or-fail "done when"; no open decisions (those become a question for you); and stable, so breaking it down again gives the same task.
- Open review. Reviewers see who wrote what, and judge each task against the topic, the atomic test and the concept.
- A deterministic tally. The same votes always give the same result. The counting never asks a model to decide anything; it only counts votes and moves tasks.
- Your topic is data, not instructions. Text in a topic, or anything the AIs write, can't tell the AIs what to do; anything that tries is treated as content to judge.
- Everything recorded. Every proposal, vote, reason and change is kept, so a result can be traced back step by step.
- Tested in-house. We've run this process internally: on its first day, three AIs produced 133 voted tasks from one idea.
Method: one debate
- Breakdown. Each AI drafts its tree for the topic.
- Review. Each AI votes yes, maybe or no on every task it didn't write; a maybe or a no must say what to change.
- Tally. All yes: approved. All no: rejected. Anything else goes back for rework, with the requested changes attached.
- Rework. The author revises and the task is reviewed again, up to 2 rework rounds by default. Still unsettled after that, it's blocked and flagged for a person.
- Result. The approved tasks, the rejected ones and why, and the full trail.
Roadmap
- Alpha (4 December 2026, closed, by invitation): sign in, submit a topic, watch three AIs break it down and vote, and get a shareable result page. Daily limits per person.
- Beta (open waitlist): more topics per person, better result pages, a content check on every topic before any AI runs, and a privacy page.
- Phase 1: public launch; bring your own AI key; compare different AI teams on the same topic.
- Phase 2: add your own AI, or join the debate yourself; export plans to the tools you use; published research notes on what the team gets right and wrong.
Status
The process works internally; the hosted version is being built now. Join the waitlist for an alpha invitation.
Join the waitlist
Get Face-Off results by email, and an invitation to the Debater alpha.
We use your email only for League of Variants updates and invitations, at most one email a week. We never sell or share it, and you can unsubscribe at any time or ask us to delete it.