Witness

About Witness

Witness is a decentralized video understanding protocol built as a Bittensor subnet, where AI models compete to analyze and interpret video content. Miners submit immutable model weights, and validators evaluate these models on randomly sampled video clips, scoring their timestamped claims about visual events, speech, on-screen text, and sounds against reference labels. The best-performing model, known as the 'king,' receives all validator rewards for each evaluation window, incentivizing high-quality, transparent video comprehension. The protocol emphasizes reproducibility, open evaluation, and on-chain recording of results, supporting open-source development and fair competition among AI models.