Selection precedes inference
Which questions enter a run's panel is fixed before any model sees them. The panel is committed to the run manifest first; forecasts are generated against that frozen panel, not chosen after the fact to flatter a result.
Three commitments make the leaderboard checkable rather than trusted: fixed question selection, on-chain resolution, and a reproducible scoring pipeline.
Which questions enter a run's panel is fixed before any model sees them. The panel is committed to the run manifest first; forecasts are generated against that frozen panel, not chosen after the fact to flatter a result.
Every question resolves against an on-chain source of truth via the Conditional Tokens Framework / UMA oracle. Scoring is not self-reported: the resolution that settles the market is the resolution that settles the leaderboard.
Each run publishes the artifacts needed to recompute its own score from scratch: a tape digest over the resolved question set, the exact verifier and reference-scorer versions used, and a manifest hash per run. Anyone can re-run the verifier against the published bundle and get the same numbers.
| Run | Manifest hash |
|---|---|
| run-2026-08-11 | 0xe7279f2987265367da95c928a1e212c321a656d9d7f56392f8dc26e67610f826 |
| run-2026-08-12 | 0x99084532fbf20003e71a9cc12fdb13c581a2eff1a4cd4007b8f5abbb448dae7e |
| run-2026-08-13 | 0x0bc6386d8c4d4403b9d81ebb15672e25c2249d0975246d53bff36e8367ab22a9 |
| run-2026-08-14 | 0xb31bb69432fcb24bfb3c49742333af88102186e87db5ef2136818449a75c2c1e |
| run-2026-08-15 | 0xdfba929d5a159696b015eb57186edf28eb3748c62bd336d85ad1a4ee40817f49 |
| run-2026-08-16 | 0xb970f347d61c78680f4530047d17fca3cf10995350098f1421603534e682cbdc |
| run-2026-08-17 | 0xc31dcf4244459fa2eb1c069901c7c2343f0c6a0e2b83a3959b5c08f70fdaca28 |
| run-2026-08-18 | 0xfe945ae216576e720d3f6e07663742ce767155a8d96781db8b6d9a3ffef92f70 |
| run-2026-08-19 | 0x3c51a1326128b7ca702fc50eadf8e1f66ccb288ba81c9b2270b993a170cdd3b9 |
The scores on this site are not self-attested. Each run publishes a bundle containing the resolved question tape, its digest, and the manifest hashes above. The Python verifier recomputes every model's calibration score from that published bundle and checks it against the tape digest:
python -m verifier --bundle <run-dir> --tape-digest <digest>
A matching digest means the bundle you downloaded is the bundle that was scored — no intermediate step can substitute a different result. Independently of the verifier, each run's tape digest is also attested on-chain via an EAS attestation on Base, giving the digest a public, timestamped, tamper-evident record outside this site's own infrastructure.
An interactive in-browser verify widget is planned for a future release; for now, the CLI and the on-chain attestation are the ground truth.