What changed
A log is a story told by the party under investigation. Today we’re opening the thing we built to replace it, and letting you license it for your own environment. Every consequential action becomes a record anyone with standing (your auditor, your insurer, your acquirer, the vendor on the other side of an incident) can re-derive without trusting us, our model provider, or our cloud.
Why it matters
In production on August 17, a coding agent proposed a destructive command against a live workload. Our first-level agent-governance gate refused it on two named predicates (wrong verb for the grant, wrong environment) and it never reached the cluster. The sealed refusal carries the grant it was checked against, every failed predicate with expected and actual, the gate version, and a snapshot hash of the exact pair: everything needed to recompute the verdict without trusting us. No score. No confidence number. No model. Just a computation over two documents.
That gate exists because we lived the opposite. A deploy agent cascade-deleted production services before anyone could step in. We wrote it up in a field note, how a SEV-0 taught us to protect our own agents, and the Evidence Graph is the record that incident should have left behind.
Availability
Available now to license for your own environment. The first five licensees take 50% off year one under the Charter program.
The full walk-through, the sealed record, and the request-access flow are on the Evidence Graph page. For the formal announcement, see the press release. It’s also what Trustworthy Autonomy™ runs on.
Known limitations
Charter pricing (50% off year one) caps at the first five licensees. After the fifth signature, Charter closes and standard terms apply.
The August 17 sealed refusal ran operator-driven through the harness gate, not a machine catching an attack on its own. The record format is what we're licensing; autonomous enforcement sits behind that.