Technical explainer · source-backed
Origin is the evidence layer for AI agents: a working prototype for agent-platform, security, trust, and infrastructure teams, currently seeking design partners. It tests proposed actions with deterministic verifiers, separates capability from permission, and produces evidence that can be independently rechecked.
Model proposes. Environment verifies. Gate decides. Trace proves. Capability is not permission.
The checked-in browser flow runs synthetic support and IAM task batteries against deterministic policy logic before production access. It produces synthetic, session-signed, unpinned evidence; execution is not attempted. Changing a bound tool or configuration makes verification fail. A valid demo signature does not establish trusted Origin issuance.
The broader design places a deterministic gate between an agent proposal and a consequential side effect. The gate would evaluate permission, scope, policy, budget, and approval requirements; only an allowed action would reach a controlled executor. This page does not claim that architecture is generally available or deployed for customers.
Origin keeps four checked-in categories distinct: machine-emitted sandbox evidence, authored examples, synthetic evaluations, and authorized fixtures. None is presented here as external customer evidence, production adoption, certification, or reviewer acceptance.
Product distinction: identity establishes who or what an actor is; observability records and debugs behavior; generic guardrails filter inputs or outputs; policy engines evaluate rules; compliance tools organize controls and evidence. Origin's intended role is to bind deterministic verification and permission decisions to a controlled action path and a re-verifiable trace.
