Frontier AI models are wrong about their own field more often than they admit. Nobody keeps a dated, public, unalterable record of it. This is that record, built as a game.
The daily loop
A contested claim about AI is picked from where the arguments happen: papers, model releases, leaderboards, release notes, the local-AI community.
The Syndicate, a panel of AI models, answers it. Seats are anonymized A/B/C/D. The lineup is recorded and hidden.
The editor, a human, reads the named public source, archives it, and writes the ruling before anything goes live. The SHA-256 hash of the ruling text is published at 6:00 AM CT. The text is not.
You blind-commit your own answer first. Then you see the panel and pick the seat you think the source backs, with a confidence from 50% to 100%.
8:00 PM CT: betting closes. The ruling text opens. Your browser hashes it and checks it against the morning's hash. The lineup is named. The entry is appended to the Ledger.
The rules that do not bend
A model never resolves a question. A human reads a named source.
Panel agreement is never called "verified." It is called "no internal evidence of error." Different claim.
A wrong answer costs you stake: your score, your streak, your grid. It never costs you access. Only conduct does.
Money never buys what effort earns.
The Ledger is append-only. Corrections are new entries on top.
Resolution terms are written before a question goes live. We resolve on what the named source says, not on "the true answer." If source and reality diverge, the source governs. That is the bet you take.
Scoring
Brier score on your confidence: if you are right at confidence p, you score (1−p)². If you are wrong, p². Zero is perfect. Confident wrongness bleeds hard; honest uncertainty bleeds little.
Known hole
You can look the answer up before you finalize. So could a Wordle player. The game is you against the Syndicate on a public streak. Cheating a free game cheats the streak.
What is not live yet
Names and track records (build step 5). Your record lives in this browser only. A second category. Player-submitted questions. These are in the spec and come after the loop proves itself.
Numbers the design rests on
When two models are both wrong they share the same wrong answer about 60% of the time. Kim et al., ICML 2025
Models favor their own generations when judging. NeurIPS 2024