Can you reconstruct what was actually knowable then?

A training table is a photograph taken now of a world that was knowable then. Features revised after the fact, labels that arrived late, records backfilled by later processes — each one leaks the future into the past, and each leak flatters a model.

The clock-dataset work builds datasets where the information state at time t can be reconstructed honestly — and measures what the gap between “recorded” and “knowable” does to evaluation.