Project Mosaic

Benchmarks skew. One prolific uploader, one content genre or one camera style ends up over- represented, and a model that happens to suit it scores well for reasons nobody intended.

Mosaic answers that with thresholds fixed in the design rather than left to judgement at ingest. One video per creator. No content domain above roughly fifteen percent of total hours. At least twelve distinct domains. At least eighty-five percent dynamic content rather than static scenes. The design specifies these as gates meant to refuse a non-conforming corpus outright; Mosaic has not been run, so no corpus has been checked against them yet.

The construction method matches Vista: the open-ended answer is locked first, the multiple- choice options are generated afterwards, and two independent blind reviews plus adjudication sit behind every item. Provenance - provider, creator, domain, licence - is recorded when material enters, not reconstructed later.