Prism - One clip, four voices, single Gemma brain

Prism follows a ground once, restyle four ways pipeline. Sample: up to 25 frames pulled across the clip in parallel, plus the soundtrack. The vision stage receives one high-resolution flow montage (the motion arc) and 15 full-resolution stills (signs, faces, on-screen text), so it sees both how the video moves and what the details say. If the clip contains speech, Gemma 3n transcribes it and the transcript joins the facts. Ground: one factual description of what actually happens - and only that. Restyle: Gemma-4-31B writes all four captions in a single structured-JSON call. Every graded word is Gemma's, in every mode; remove one API key and Prism runs pure-Gemma end to end. The demo showcases the wider Gemma family: EmbeddingGemma scores each caption's semantic anchor to the grounded facts, the language selector transcreates all four voices into 16 languages with tone intact, and the listen button speaks with T5Gemma-TTS - a Gemma voice synthesized on AMD silicon (Radeon W7900). We measured everything and published it. GEMMA_FINDINGS.md in the repo documents Gemma-4's OCR limits (9px text read at 512px input), parallel scaling (six calls in about a second), payload ceilings, and A/B results from the live judge - including the changes that scored worse. An honest model-role table shows exactly what each model does. Reliability is the floor: results are pre-seeded and rewritten atomically per clip behind a three-tier Gemma failover, so Prism never returns nothing.