Signal, Timing, or Architecture? Learning from Time Series That Are Not Yet Final

Operational time series are not static: observations arrive late, are revised for months, and are routinely evaluated against values unavailable when the forecast was made. We formulate forecasting over versioned release events with separate reference and availability times, and build an event forecaster over both clocks that reads the histories of a series and its neighbors with their revisions, predicts a residual over persistence, and admits context through a single removable gate. We ask whether an apparent gain arises from the observed values, from their association with release times, or from the model architecture. We formalize input interventions on a fixed training procedure as two controls at matched capacity that preserve the release schedule: a null that removes values, and a shuffle that breaks only their pairing with releases. The two test different hypotheses and can disagree, on one archive in sign. In sealed evaluations registered in advance and opened once, authentic release content improves hospitalization backfill forecasts over both controls, while no gain is detected on a macroeconomic vintage archive. Development comparisons show that a content gain need not beat persistence or archived expert forecasts. In a separate synthetic study, disabling the continuous flow of a model for irregularly sampled data leaves most of its apparent advantage over a conventional baseline intact. These results separate what a model uses from how well it forecasts, and argue that evaluation of delayed, revised time series must do the same.