Calibration record
The public ledger
Every prediction SONAR locks before an election is listed here with its content hash and timestamp. When the result certifies, the lock is scored against it. Nothing is added after the fact, and nothing is removed.
Misses publish at the same volume as hits. A method that only shows its wins is not a record. This page is the whole record, scored and pending together.
Ledger exported loading from the production database. Hashes match the append-only lock log.
Loading the ledger.
How to read a row
- Locked prediction. The two-way normalized share SONAR published before the election, frozen with a content hash.
- Certified result. The official result in the same two-way normalized space. Shown once the result is in.
- MAE. Mean absolute error in share points between the locked prediction and the certified result, averaged across the candidates.
- Winner called. Whether the locked prediction named the eventual winner. A correct winner with a wide share error is a called winner and a real error, and both are shown.
- Benchmark. The public poll on file at lock time, where one existed, so SONAR can be scored against it rather than in isolation.
The South Carolina pair is scored: SONAR named the winner, with a share error of 12.8 points on a low-attack runoff. The Arizona lock is standing and will be scored after the July 21 primary certifies. It carries a public polling benchmark, so its score will report both the SONAR error and the polling error side by side.
Why lock at all
A prediction made after a result is not a prediction. Locking a content hash before the election, and committing it to an append-only log, is what lets a later score be trusted. Confidence in any SONAR standing comes from this record and from pool volume, and from nothing else.