High-resolution convection-permitting models (CPMs) provide vital datasets for studying localized extreme precipitation and associated multi-hazard risks like landslides and floods. Yet validating CPM datasets against observations remains a challenge. This study evaluates two sub-daily CPM datasets, VHR-REA_IT and LAM-HIND, against observations over a 10-year period across two Italian regions: Emilia-Romagna and Campania. Extreme precipitation is defined using the 95th percentile of wet-event intensities across accumulation durations of 1, 2, 4, 12, and 24 hours.</p> <p>Conventional verification metrics only partially assess dataset performance and hide differences in the temporal organization of precipitation extremes. Therefore, a new nested time-window framework is presented to better capture the association between short- and long-duration time window extremes (STWEs and LTWEs, respectively). Observations show that STWEs and LTWEs are only partially associated. Depending on region and season, around 7–15 % of observed 1-hour STWEs are embedded within 24-hour LTWEs. CPMs reproduce this cross-timescale organisation substantially better than ERA5, yet are overestimated, indicating excess temporal persistence of simulated heavy-rainfall. Consequently, LTWEs are more dependable than STWEs, as CPMs distribute short extreme precipitation bursts into longer durations with persistently higher volumes.</p> <p>Differences between the CPM datasets are small. LAM-HIND exhibits subtly higher detection and event-overlap scores but a higher false-alarm rate and stronger cross-timescale coupling overestimation. VHR-REA_IT more closely reproduces the observed temporal association between short- and long-duration extremes. These differences may stem from the ‘double penalty’ effect, choice of an intermediate nesting strategy, and sensitivity in resolving sub-grid turbulent fluxes. However, they are both substantially improved over coarse products like ERA5 for sub-daily extreme precipitation.