Multimodal safety scoring on children's video — transcript, frames, and family values aligned.
SafeWatch-Bench evaluates how well a multimodal system scores children's video content for safety. Inputs include speech transcripts, key frames, and metadata; outputs are calibrated 0–100 scores with explainable sub-dimensions parents can configure.
This eval powers Lemu Kids — every content score in production is regression-tested against SafeWatch-Bench before release.