Lab · Runs entirely in your browser
Casting Vote
Two AI models run on your own GPU and argue a motion, talk a question through, or improvise a scene. You chair the sitting, and you hold the deciding vote.
Order paper
Qwen3 reasons before it speaks. It takes longer, and you can read its reasoning on every card: test-time compute you can watch.
The performers must accept every offer and add to it. No blocking, no stepping out of character. Shout a new suggestion any time.
The first sitting downloads the models once; after that they load from your browser's cache. Nothing you type leaves this tab.
The motion
How a sitting runs
- Each round is written blind. Neither model sees the other's speech for that round until both are in.
- You chair. Between rounds, put a question to either seat, stop a speaker mid-sentence, or call the vote early.
- The division. Both models vote honestly, sealed so they can't sway you. Then you cast the deciding vote.
Why two models? One model answering alone brings one model's blind spots. Here two different models spend extra thinking on the same question, each pushed by what the other just said, and you are the outside influence: question them as it runs, then decide. It is test-time compute with a human in the loop.
Yes, and… turns the chamber into an improv stage: give a suggestion, and the two models build a scene line by line, accepting every offer, until you call "Scene!" Then you decide who carried it.
The division
You · the chair
The seats are voting…
The reading
A decision model (strands-decider-2B) reads the speeches, without the votes or the models' names, and gives a probability for which side made the stronger case. It decides nothing. Your vote is the ruling; the reading appears only after you cast it, so it cannot sway you. When it is not sure it says "Too close to call".
The record
Rulings you make are kept here, in this browser only.