Most AI tools are optimized to agree with you. Ask one to poke holes in your idea and it'll find a couple of soft ones, tell you it's "promising," and hand you back your own confidence with better formatting. That's the problem, not that AI can't reason, but that it has no incentive to actually try to break what you give it. The inspiration was pretty direct: I wanted something that treated my own claims the way a hostile reviewer would, not the way a cheerleader would. Bring a conjecture, a startup thesis, a systems design, whatever, and have it attack the specific thing that field's own experts would attack, degenerate cases for a proof, unit economics for a thesis, tail latency for a systems design. Not generic "have you considered scalability" feedback. The real thing. The approach changed a lot once it was actually being used against real claims instead of toy examples. Early on, confidence scores came straight from the model, which meant the system judging a claim and the system making it were the same thing, worthless. That got replaced with a deterministic evidence-ranking layer: a claim can only earn a high confidence number if something outside the model actually checked it (a numeric kernel, a Monte Carlo sweep, a SAT search), not because it sounded plausible. The other big shift was realizing corrections needed the same scrutiny as everything else. It's easy to build a tool that folds the moment a user pushes back. Getting it right meant a bare "you're wrong" gets sent back for a reason, and even a real device-side refutation stays on the record as refuted, not quietly erased. That's the part that actually took the most iteration, since it's the one place where being technically correct and being usable pull against each other.

Panton

Truth isn't generated. It's stress-tested.

Truth isn't generated. It's stress-tested.