NeuPortal
AI forecasts with a public, Bitcoin-timestamped track record
Every AI product on your feed claims it can predict something. Almost none can prove a single prediction was made in advance. So we built our product around the opposite promise: make cheating impossible, publish everything, let the record speak.
The protocol: before every World Cup match, our model's probabilities are locked into a file and hashed into Bitcoin via OpenTimestamps (free, nobody can backdate — including us). At the same instant we freeze Polymarket's price as the benchmark. After the result, both get a Brier score. Everything goes on a public board. There is no delete button.
11 matches in, the honest scoreboard: the market is closer on 7 nights out of 11. Our model leads on average error by a margin that's statistically meaningless at this sample size — and we say so on the board.
Why publish numbers where we're losing? Because "verifiable but modest" beats "impressive but unfalsifiable" — and because by the time the sample is big enough to mean something, the record will already exist with zero curation opportunity.
Live board: https://neuportal.ai/experiment
Raw stamped files + verify instructions: https://github.com/neuportalai-ui/experiment-proofs
Would love critique of the protocol — especially holes that would let us cheat that we haven't closed.
1 Comment
The thing that stood out to me isn't the predictions—it's the willingness to make being wrong part of the product.
Most AI products optimize for looking accurate. You're optimizing for making accuracy auditable. If users start trusting the process more than any individual prediction, that becomes a much harder position to copy than simply claiming a better model.