Can You Prove When an AI Model Was Released?

A model's release date and knowledge cutoff come up in surprisingly high-stakes contexts: legal disputes over who published something first, journalists checking whether a chatbot could plausibly have known about an event, researchers citing "the version available as of a given date." In all of these, the date is usually taken on faith -- from a blog post, a Hugging Face "last modified" timestamp, or a README.

Why those sources are weaker than they look

Blog posts can be edited after publication. Hosting platform timestamps reflect whenever a file was last uploaded, which can change if a repo is refreshed, and platforms themselves don't promise the historical timestamp is tamper-proof. None of this is usually malicious, but none of it is independently verifiable either -- you're trusting the platform and the publisher.

What an on-chain attestation actually proves

Writing a hash and a claimed date to a public blockchain, in a non-revocable record, creates something meaningfully different: a timestamp that would require rewriting the blockchain's history to alter -- specifically the kind of tampering these systems are built to make impractical. Anyone can look up the record independently, without asking the model's publisher or the registry that submitted it.

The honest caveat

This proves a specific claim (this hash, this name, this claimed cutoff date) existed and was attested at a specific time -- not that the claim itself is true. If someone submits a false claimed cutoff date, the attestation faithfully proves that false claim was made at that time; it doesn't independently verify the model's actual training data cutoff. What it does rule out is the claim being quietly backdated or edited later, which is the failure mode a blog post or file timestamp can't rule out.

Weight Registry's Timeline shows real examples of this across dozens of open-weight releases.