The record starts at zero
No version has yet been published through the verification gate, so there is nothing to measure. We show that as it is: every count is zero and no rate is shown, because a rate over nothing would say nothing true. The figures appear here as the gate does its work.
Figures are recomputed every Monday and cover everything up to 9 November 2026.
Earlier versions
252 versions were published before multi-model verification, by an earlier version of Juriset after a person reviewed each one. They are in the API with their full history, but they are not in the figures above, because they were not checked the way new versions are. The earlier system recorded 34 corrections or rollbacks among them.
How each figure is computed
- What is counted
- Versions published through the verification gate: a model from one vendor extracted the rule, a model from another vendor re-extracted it blind from the source and they agreed on every required field, or a person approved the escalation. Versions imported from Juriset V1 are not counted.
- Published automatically
- Versions published with no person involved ÷ all published versions.
- Sent to a person
- Versions the verifier checked that were sent to a person ÷ all versions the verifier checked.
- Corrected
- Published versions later corrected, rolled back or withdrawn ÷ all published versions, overall, by publisher (auto or human) and by rule type.
- Sample audits
- Weekly sample audits that matched the source ÷ audits decided. 5% of auto-published versions are sampled each week (10% for higher-stakes rule types).
- Time to resolve
- Median hours from the gate sending a version to a person to that person approving or rejecting it.
The same figures, in machine-readable form, are at GET /v1/track-record.