The problem with storing prompts

The most common question we get from engineers integrating sentinel: "we can't store our prompts — they contain personal data. does that mean we can't use audit?" The answer is no. The reason is zero-knowledge proofs.

What a zk proof actually is

A zero-knowledge proof lets you prove a statement is true without revealing the information that makes it true. In the context of llm audit, the statement is: "this model received this prompt and produced this response, and neither has been modified." The information not revealed is the content of the prompt and response themselves.

How sentinel implements it

Sentinel uses a zk-snark over the seal hash. When you call sentinel.prove(sealHash), the sdk constructs a proof that: (1) you hold the pre-image of the hash — meaning you have the original content, and (2) the hash in the ledger matches. The proof is approximately 200 bytes. It can be verified by anyone with the public verification key in under 10ms, without them ever seeing the prompt or response.

What this means in practice

A healthcare company can prove to a regulator that a specific model interaction occurred, that the model version was what they claimed, and that the log was not modified — without handing over patient data. A legal team can prove the model produced a specific output without disclosing client communications. The proof is the audit trail, and the proof contains no content.