Scored against what actually happened

Where a realized fill of comparable size is subsequently observable on the same venue, the estimate is scored against it. What gets published is the error distribution — by size bucket, by venue, by market condition — not a single vanity accuracy percentage.

A headline number like “97% accurate” is unfalsifiable and hides the only cases anyone cares about. The distribution is harder to read and impossible to fake, which is the point.

This is deliberately uncomfortable

It means publishing the cases where the estimate was wrong, including the tail. The alternative — asking consumers to trust an unaudited number — is precisely the failure mode Crifine was built to attack. A product whose thesis is “the number you are trusting is not checkable” cannot itself ship an uncheckable number.

The design constraint

Scoring has to be automatic and adversarially reproducible from published data. If Crifine is the only party able to compute its own accuracy, the score is worthless. Every input to the score is therefore published on the evidence pages the score derives from, so anyone can recompute it and disagree.