Test the temporal evidence boundary before interpreting quality

Choose a group based on the intended deployment question: episode, source recording, camera/session, participant, or another justified unit. GroupKFold's official contract keeps a supplied group out of both sides of a fold, but choosing group labels correctly remains the experimenter's responsibility. Hash-based duplicate checks, provenance links, and near-duplicate review can find re-encodes or extracts with different filenames. Group separation does not prove every semantic duplicate is absent, so state what identities and similarity checks were actually covered.

For a chronological holdout, inspect the full feature-support interval, not only the target or decision timestamp. A validation clip with twenty seconds of lookback can overlap training evidence even when its decision occurs after the split boundary. Purge or otherwise handle that overlap under the declared task. Some deployment tasks legitimately use earlier observed history, but their evaluation must explicitly reproduce that allowed operational context. Reusing hidden training state or fitting preprocessing on held-out inputs is a different information policy.

Before scoring a model, test timestamp/timebase conversion, clip endpoints, dropped-frame handling, image orientation, and label-coordinate mapping on known fixtures. Include an irregularly timed source so a frame-count shortcut fails visibly. Compare predicted event intervals with the declared overlap and matching rules; a correct class name does not guarantee correct localization. ActivityNet's official evaluator combines temporal overlap, confidence ordering, and instance matching, so a bare overlap value should not be called a complete average-precision result.

Class imbalance also changes what a high accuracy means. Report the class counts and suitable per-class recall in addition to any aggregate. Balanced accuracy averages class recall over the declared nonempty classes, giving rare and common classes equal weight under that convention. It does not automatically address label noise, temporal dependence, or deployment prevalence. Declare threshold selection and matching before looking at final held-out outcomes, then keep the same cohort and denominators when comparing model versions.

A release gate can combine provenance checks, a source-group split audit, timestamp/label fixtures, and task metrics with predeclared tolerances. Investigate failures by category instead of hiding them behind an average: temporal shifts, missing events, minority classes, and encoder/index mismatches have different mechanisms. Retain the exact split and preprocessing manifests so a later improvement can be reproduced. Reusing the held-out set repeatedly for selection also changes the interpretation of an untouched-test claim.

The sampling lab shows why an independence assumption matters, without verifying it for videos. Reset, set both p=0.5,n=20, predict mean-proportion variance 0.0125, then Play or Step. Reset and set both p=0.5,n=100 predicts variance 0.0025. Its masses describe independent Bernoulli trials; playback highlights outcomes rather than drawing samples. Correlated frames are not thereby independent observations, and the displayed standardized-CDF bound does not certify individual bars or a real video's uncertainty. Group-aware uncertainty needs an appropriate sampling design.