Release notes · 15 November 2026
Chestertown 0.1.0
Chestertown is an open, open-source architecture for governable, enforceable and verifiable AI systems. Prudence, AOS and Ernest are its reference implementations, and they close one loop: decide → run → prove → oversee.
| App | Role | Version |
|---|---|---|
| AOS | Run: declared, deterministic pipelines in YAML, with human steps, RBAC and protocol steps (MCP, A2A, gRPC); natural language only picks and fills a pipeline | 0.5.0 |
| Prudence | Decide and Oversee: AI system registry with EU AI Act risk tiers, policy decision point, approvals, oversight, incidents | 0.1.0 |
| Ernest | Prove: tamper-evident evidence, anchoring, offline-verifiable receipts | 0.3.0 |
Each app keeps its own version and can run on its own. All three are Apache-2.0.
The loop
- A pipeline in AOS that belongs to an AI system asks Prudence before it runs. Prudence answers
ALLOW,DENYorAPPROVAL_REQUIREDfrom the system's approval state and the active policy; AOS enforces it and fails closed if Prudence is unreachable. - The run's evidence travels to Ernest over Kafka, and Prudence's decisions and reviews through a durable outbox. Pipelines never wait for evidence.
- Ernest chains, batches and anchors the evidence (OpenTimestamps, EVM, RFC 3161), and anyone can verify a receipt offline, in the browser or the CLI.
- Prudence watches the estate's posture and drift, and the incidents and periodic reviews that send a system back for approval.
Beyond AOS: eighteen agent frameworks
Any runtime can take part in the loop. Prudence ships adapters for eighteen agent frameworks, all on one contract: twelve in Python (LangGraph, LangChain, CrewAI, LlamaIndex, OpenAI Agents, Claude Agent SDK, Google ADK, Pydantic AI, smolagents, Agno, Strands Agents and Microsoft Agent Framework), four in TypeScript (Vercel AI SDK, OpenAI Agents JS, Mastra and LangGraph.js) and two in Java (Spring AI and LangChain4j). Each asks the Decision API before a tool runs, reports the execution and seals its evidence in Ernest under the same id, so Prudence matches every reported run with its proof. A conformance kit checks any adapter against the contract, a conformance suite runs each one against a real agent, and the demo runs the Python ones from Prudence's console.
Agents without an adapter can be governed by an AI gateway instead. Prudence ships a LiteLLM guardrail that asks the Decision API before every model call and tool call, with the agent's own token, and lets a call through only when the answer is an ALLOW about that very agent. Prudence records who asked (the gateway), which agent acted and, when a verified token says so, for which person; each AI system declares its agents, and a gateway may speak only for the agents it is allowed to. The same twelve Python example agents run behind it unchanged.
Governance functions: one authorization model across the three
Chestertown does not hard-code job titles. It defines twelve governance functions:
system.declare · system.approve · policy.author · policy.approve · oversight · incident.handle · ai.use:<domain> · evidence.write · audit · platform.admin · runtime.decide · runtime.report
An organization writes one governance model that maps its identity provider's roles or groups to those functions and sets its rules: how many people approve each risk tier, which functions one person may not combine on the same system or task, and which accountabilities a system needs. The model can tighten the rules but never go below the legal floor: a high-risk system always needs two distinct approvers and a supervisor (AI Act art. 14).
- Every app resolves the same token to the same functions, checked by shared test vectors in Python, Java, Go and TypeScript.
- Templates for a small or medium organization, a bank (three lines of defence) and a public body.
ct-modelchecks a model for gaps and generates a Keycloak development realm, Entra ID app roles and app settings.- Validated against Keycloak, Dex and Microsoft Entra ID with the same acceptance check.
Observable end to end
The three apps follow one set of observability conventions, so a governed run can be followed across them:
- OpenTelemetry tracing (OTLP, W3C Trace Context): AOS's call to Prudence continues the same trace; evidence delivery and processing are new traces linked to the run, never held open.
- One business id per flow: the AOS task id travels as
correlationIdthrough the Decision API, the outbox, Kafka and into Ernest's evidence, so a run, its decision and its proof are found together, days later. - JSON logs with the same fields in every app, each line about a run carrying its trace.
- RED, runtime and domain metrics (decisions by verdict, gate latency, evidence delivery delay, chaining delay, proof time), with exemplars from a latency bucket to its trace.
- Service-level objectives: 99.5 % over 28 days for admission, evidence durability, verifiability and oversight on complete data, with multi-window burn-rate alerts and Grafana dashboards; tail sampling keeps every error and every governed flow.
- No secrets, prompts, parameters or personal data in any signal.
Contracts between the apps
The Decision API, the decision events and AOS's audit events are versioned contracts with test vectors that both sides run (the copies carry SHA-256 sums), and a nightly end-to-end suite runs AOS, Prudence and Ernest together with Keycloak and Kafka: identity, verdicts, evidence, failure modes and observability.
Supply chain and security
Release images are signed with Sigstore cosign and carry an SBOM and build provenance; a critical vulnerability with a fix blocks publication. Every repository runs CodeQL, a secret scan of its whole history and Dependabot. Each app has a threat model and a private vulnerability reporting channel. The consoles meet WCAG 2.1 AA, checked with axe in their browser tests.
Milestones
- October 2025: Ernest starts.
- November 2025: AOS starts.
- June 2026: Ernest v0.1.0-alpha, its first public release.
- July 2026: AOS v0.3.0 as a Go module, and Ernest v0.2.0-alpha.
- October 2026: Prudence starts; the three projects become Chestertown.
- November 2026: Chestertown 0.1.0.
See it
ccmlab.cc has a recording of each case, one per risk tier, made on the live demo. Live sessions of the demo open from January 2027. Each app's README explains how to run it on its own; the infrastructure (Kafka, MongoDB, PostgreSQL, OPA, an OIDC provider) is yours to provide.
Status and limits
- Prudence is a governance-readiness platform, not a legal declaration of EU AI Act conformity.
- Ernest is not a qualified trust service provider under eIDAS, nor a qualified electronic ledger. It can rely on one: an RFC 3161 timestamp from a qualified provider carries the legal presumption of accurate time and integrity; OpenTimestamps and EVM anchors are technical evidence anyone can verify, without that presumption.
- All three are 0.x versions. Ernest publishes its criteria for beta, and two are still open: a pilot with a security review, and the data lifecycle (designed). AOS and Prudence are the first public releases of their current shape.
- The consoles were checked against Keycloak and Dex; against Entra ID only the APIs were.
- These are reference implementations, maintained by one person on a best-effort basis: issues and pull requests are welcome, but there is no support and no promised response time.
Author
Developed and maintained by Carlos Castro Martos. This project was developed independently. It is not sponsored, endorsed, or licensed by any employer, organization, or company.