Align clip support, event labels, and the prediction cutoff

Use one timebase and origin for the clip and annotation, then specify endpoints. Half-open intervals [s,e) avoid counting one boundary instant twice when adjacent segments meet. The overlap duration is the positive part of the earlier end minus the later start. A clip-positive rule might require any overlap, a minimum fraction of clip duration, a center point inside the event, or another task-specific criterion. These rules can produce different targets for the same annotation and must not be silently mixed.

Localization evaluation asks another question: how closely does a predicted event interval match the annotated interval? Temporal intersection-over-union compares intersection with union, while benchmark average precision additionally depends on confidence ordering, class, and matching conventions. Do not relabel a clip's overlap fraction as temporal IoU or infer benchmark AP from one interval. Preserve the original event identity and uncertainty when annotations are approximate, so preprocessing does not create artificial millisecond precision from coarse human labels.

For an online decision at a declared cutoff, input features must satisfy the task's causal availability rule. A future outcome can be a legitimate training target, such as an action beginning in the next second, while future images must not be included as already observed inputs. Offline annotations can be created later to identify that outcome. Their later creation is not inherently leakage; using annotation-derived future information as a model input would be. Distinguish label knowledge during training from what the deployed predictor may observe.

Presentation time alone is insufficient for availability because decoding can require reordered references and transport can delay frames. Define a mapping between media position and the observed decision clock, then record when each feature becomes usable. An offline clip model that includes centered context or bidirectional temporal attention is not automatically an online model. Evaluation should either allow the stated lookahead and count its latency or enforce a causal feature construction and compare against that different contract.

Retain source video and event IDs, source intervals, chosen frame PTS values, target cutoff and horizon, label rule, and missing-data policy. Include clips that straddle action boundaries, contain multiple actions, or have no valid sampled frame in an intended subinterval. Check whether a clip was selected using label information in a way that changes the evaluated population. Selecting only obvious positive moments can be a valid training policy but does not represent an unfiltered online stream without a corresponding evaluation design.

The probability lab isolates a hypothetical rate of clips satisfying a declared rule. Reset, set both p=0.5 and n=20, predict ten expected qualifying clips, then Play or Step. Reset and set both p=0.75,n=20 for fifteen. These exact binomial counts assume independent identical qualifications and are not sampled video clips. The instrument computes no interval overlap, event label, feature-ready time, or temporal correlation; it illustrates why a rule's selected population should be counted rather than treated as the whole source stream.