What each source actually contributes

• captured — records we hold for this model in the window, as received
• unique — distinct (model, task, site, target time, horizon); repeats of the same case collapse
• matched — a second model answered the identical case, so a comparison is possible at all
• eligible for the objective — matched, with a stated forecast and a matured outcome
• exported — eligible and inside the frozen delivery window

• Objective named for the "eligible" stage: relative selection on shared cases: choosing between models on a case both answered, scored against a matured outcome. A record that fails it may still serve another objective; the objectives column says which.
• Our native engine produced no receipts in this window, which is why it contributes zero shared cases: the engine's outputs are stored as its own forecasts and scores, not as receipts against the same sites, tasks and target times the reference roster answers. Until it runs on the same upcoming opportunities as that roster, no shared case can exist, and we do not manufacture one retrospectively.
• Notflat evidence is already represented here through receipts: 2 model lanes arrive as receipts and are counted in the stages above. The mirrored snapshot tables hold the same provider records natively and are not counted again, so the two are never summed.
• "Consumed by Zeno" is deliberately empty. We do not know what a consumer trained on; the column is filled only when that accounting is supplied to us.
• Single-model histories are retained in full. They are unsuitable for relative-selection labels only, and remain usable for signed error, bias and calibration.

Computed 2026-09-23 23:08 UTC by the maintenance cycle.