Conformal Coverage Guarantees for Any Video Temporal Grounder

Cover is a post-hoc, model-agnostic wrapper for video temporal grounding: instead of returning one unqualified interval, it calibrates a temporal nonconformity score on held-out labeled examples and widens each prediction into a region with finite-sample, distribution-free coverage of at least 1 − α under exchangeability. It requires no retraining or white-box access, and works with trained localizers such as QD-DETR and black-box video-language models such as Qwen2.5-VL-7B-Instruct. The method supports a length-normalized two-sided boundary score, a separate per-boundary calibration for asymmetric errors, and a super-level-set score that can produce disconnected relevance regions. Across Charades-STA, ActivityNet-Captions, and QVHighlights, covering three benchmarks and five grounders, realized coverage closely follows the requested target: for example, ActivityNet-Captions reaches 0.895 at a 0.90 target. Cover also exposes why learned uncertainty is not automatically reliable: the evidential SRAM grounder achieves only 0.660 against its nominal 0.80 target, whereas Cover restores calibrated coverage on the same predictions. Fixed, hand-picked margins are unstable, producing coverage from 0.096 to 1.000, a 0.904 range across models and datasets. Calibration also reveals that Qwen’s boundary errors are concentrated at event offsets, while its onsets are nearly accurate. The main limitation is that guarantees are marginal and depend on exchangeability; dataset shift, within-video dependence, or conditioning on difficult subgroups can reduce validity, motivating weighted or Mondrian calibration.