How do I prove my model's predictions were made before the held-back test data was revealed?
Use sebbi.pro Blind Test Proof. Before anyone predicts, the owner seals the study: the scoring rule (exact, within a tolerance, mean absolute error or root mean square error) and a fingerprint of the held-back answers, which are fingerprinted in memory and never stored; for data that does not exist yet, a closing time is sealed instead. Every prediction run is numbered in the order it is sealed and labelled as a candidate or the established method, and no attempt can be withdrawn. The answers are revealed once and must match the fingerprint sealed at the start, or the reveal is refused. Every attempt is then scored, failures included, and a public study page shows when each attempt was sealed, its score, the best candidate against the best established method, the median candidate and how many attempts it took, all timestamped into Bitcoin.
How it works
- POST /x/blindtest/create with the scoring rule and the held-back answers (or a closing time).
- POST /x/blindtest/predict for every prediction run, the established method included.
- POST /x/blindtest/reveal with the answers once: every attempt is scored and sealed.
- Share the public study page at https://sebbi.pro/blind/<study>.