Measurement Availability and the Time Course of Transient-Stability Classification in a Public 39-Bus Benchmark

A transient-stability classifier is only meaningful for a specified measurement horizon. We examine a publicly released New England 39-bus benchmark by linking 3,120 labelled feature records to 9,360 archived simulated trajectories. Numerical matching identifies 147 early-coded inputs at 1.000 s, but 146 later-coded features, 60 unsuffixed generator channels and one irregular voltage feature at later recorded samples. Conditioning on retrospectively recovered short-circuit type reduces the minimum observed-file error count among identical early input groups from 453 to 32. In five-fold exact-state-disjoint validation, balanced accuracy rises from 0.5360 with the early view to 0.9847 with all 354 features; the latter includes later observations and is not an early-warning result. We additionally train time-indexed models on 2,928 records uniquely linked to archived events, using only generator terminal voltages observed at the nominated absolute simulation time and their change from 1.000 s. Balanced accuracy is 0.5253 at 1.000 s, 0.9330 at 1.050 s, 0.9743 at 1.150 s and 0.9837 at 1.433 s. Mean leave-one-load-level-out balanced accuracy for voltage plus rotor speed is 0.5279 at 1.000 s and 0.9856 at 1.433 s. These results quantify an observation-time-dependent change in discrimination on a fixed recorded cohort.