This page describes the shape of the approach. The detailed methodology note — estimation design, validation protocol, and the full explainability specification — is shared privately with methodologists under agreement, because it is the level of detail your staff will want to argue with.
Validation
Validation reporting is produced per deployment, against the office's own ground truth — a round it has already collected and published. That is the point of reconstructing rather than predicting: the accuracy of the method is a fact your office can establish for itself, on its own data, before relying on it.
We publish no accuracy figures on this site. A number produced on someone else's data, on a different instrument, in a different country, would tell you nothing about your own — and we would rather say that than quote one.
Inputs, configuration, model version, uncertainty, and explanation travel with each estimate. A figure you cannot trace is a figure you cannot defend in public, and official statistics are defended in public for years.
Explanations are produced at model and record level as a standard deliverable. They exist so your methodologists can disagree with the platform on substantive grounds.
Output is a ranked list of divergences for analyst review. No estimate is published, substituted, or acted on without a decision by your staff, and every decision is logged.
The approach is documented to map onto the standard statistical business process stages and the international principles governing official statistics.
Stated limits
The full note — estimation design, validation protocol, explainability specification — is shared with methodologists under agreement.