Four roles, and the open contracts between them.
The architecture
Decision, enforcement,
evidence.
Chestertown defines how a governance decision becomes enforcement, and how enforcement becomes evidence anyone can verify. It is three things, in this order: an architecture, the implementations that prove it works, and a deployment that runs them.
Prudence, AOS and Ernest: open source, 0.x, and each usable alone.
The live demo: all three running as one, with one sign-in, shown live from January 2027.
Four roles
No verdict, no run.
- DecideThe policy decision point answers allow, deny or approval required. The policy lives here, outside the runtime. Implemented by Prudence.
- RunThe enforcement point asks before it acts and obeys: AOS before each run of an AI pipeline, the Prudence adapters before each tool call in an agent framework. An error or a timeout counts as a no.
- ProveThe evidence store keeps decisions and executions so they can be checked independently. Implemented by Ernest.
- OverseePosture, drift, incidents and expiring approvals feed the next decision. Implemented by Prudence.
Open contracts
The contract outlives
the implementation.
Each contract says what a role must do, and most ship test vectors that both sides run. An implementation conforms when it passes them, not because of where it came from.
Decision API v1
How a runtime asks whether a step may run, and reports what it ran. OpenAPI plus test vectors for both sides.
OpenID AuthZEN 1.0
A standard way in for MCP gateways and generic authorization clients. Anything but ALLOW answers false.
Governance functions v1
A fixed vocabulary of who may do what, mapped to each organization's identity provider.
Evidence v1
What the decision point records as evidence, and what it reads back to confirm it.
Decision events v1
Each decision as an event, for delivering evidence at volume over Kafka.
Audit events v2
How a runtime's audit journal reaches the evidence store, independently of the decision point.
How it fits what you already run
How it fits what you already run
- Your runtime
- Keep the framework you use. LangChain, LangGraph, CrewAI, Google ADK, the OpenAI Agents SDK and others have adapters; for anything else, the Decision API is two HTTP calls. AOS is a reference, not a requirement.
- Your identity
- Sign-in and machine identity come from your OIDC provider, and your groups map to the governance functions.
- Your infrastructure
- It runs on standard pieces you operate: PostgreSQL, OPA, MongoDB, Kafka, OpenTelemetry. The implementations need them; they do not hide them.
What it does not claim
What it does not claim
- Early
- The implementations are 0.x. Prudence is the only decision point so far, and the contracts still live next to their implementations.
- Tamper-evident
- Ernest makes changes detectable, and anchoring limits what an operator could rewrite. It does not make a database immutable. Object-lock storage keeps records; Ernest makes them independently verifiable. A plain external store gives custody, not proof.
- Scope
- Only what passes through an enforcement point is governed. It helps put the EU AI Act into practice; it does not make anyone compliant.
The architecture in one line
Decide, run, prove, oversee.
In that order, every time.