What this is
A council of frontier AI models argues against each other to stress-test one yes/no decision. Critics attack every argument across multiple rounds; weak arguments are pruned. What survives is your verdict.
Fits
Decisions where being wrong is expensive -- committing capital, hiring, picking between options with hidden trade-offs, anything you can't easily reverse.
Doesn't fit
Chat, open-ended questions, factual lookups, advice you'd get from a quick web search.
Try one
Progress
Waiting for first round...
May converge faster than the estimate if the council agrees early.
Debate
Cost so far: $0.0000
Verdict
Motion:
Narrative
Surviving arguments (0)
Pruned arguments (0)
Technical details
Stop reason:
Rounds completed:
How to read this verdict
Verdict types
- APPROVE -- arguments for the motion survived cross-examination at high confidence.
- REJECT -- arguments against survived; the motion didn't hold up.
- INCONCLUSIVE -- every argument was pruned, or survivors are split / low-confidence.
- STOPPED -- the safety layer flagged a possible prompt-injection attempt and stopped the debate before a verdict was reached.
Confidence (0.00 to 1.00)
Aggregates the surviving arguments, weighted by stance. 1.00 means survivors agreed strongly; 0.00 means even split or no survivors at all.
Surviving vs pruned arguments
Each council member submits arguments per round. Critics attack them. An argument is pruned when its own author's revised confidence drops below the survival floor under critique pressure. What's left is what stood up.
Stance badges
Each argument is FOR the motion, AGAINST it, or NEUTRAL. A confident AGAINST survivor pushes the recommendation toward reject; a confident FOR survivor pushes toward approve.