The notary
Evidence Runtime
A signed, replayable record of every AI decision — so what happened is an account, not a claim.
The Evidence Runtime turns each AI decision into a signed, hash-chained record. Inputs and outputs are hashed; each record links to the one before; a signature closes the chain. Change any field and re-verification fails at exactly that link — deterministically.
The model stays probabilistic. What becomes deterministic is the record of what the model did and what the runtime decided about it. The record, not the model’s word, is the account.
The artifact it produces
A signed, replayable per-decision record, emitted in AIREP — verifiable offline by anyone, tied to no vendor.
What it does not claim
It does not make the model correct. It makes what the model did provable. A wrong-but-signed decision is still wrong — it is just no longer deniable.
Where it fits
- Regulated workflows where an audit trail is non-negotiable
- Multi-agent systems where one agent’s claim feeds another
- Coding agents that assert "fixed / tested / deployed"