Chestertown · an open architecture

AI that can
answer for itself.

Chestertown is an open, open-source architecture for governable, enforceable and verifiable AI systems. Prudence decides whether a system may run, AOS shows how a runtime obeys, and Ernest keeps a record where tampering shows. Built around the EU AI Act, deployed, and recorded running.

See it run · 1:21

One request.
Denied, approved, proven.

An agent asks to look up a customer's data. Prudence says no: nobody has approved its AI system. A different person approves it in Prudence, the same request runs in AOS, and Ernest keeps the evidence.

Recorded on the live demo, 7 Nov 2026. Nothing staged: the real apps and buttons. No sound; captions on screen.

Transcript
  1. This agent wants to look up a customer's data. Nobody has approved its AI system yet.
  2. Before every run, AOS asks Prudence. Customer-support lookup, limited risk.
  3. Denied by Prudence. AOS did not run a single step: fail-closed.
  4. Prudence, AI governance. The system owner submits it for review.
  5. A different person reviews it. Limited risk needs one independent sign-off.
  6. Approved. The approval itself is recorded as evidence.
  7. Same request, again. Nothing changed in AOS: only the governance decision.
  8. Prudence answers ALLOW. AOS runs the read-only lookup and changes no data.
  9. Ernest, evidence ledger. Prudence's verdict and every AOS step, each with its own hash.
  10. The chain is verified. Change one record and the check fails.
01

Make responsibility explicit before an AI system reaches people.

02

Keep an evidence trail that can be inspected, not merely asserted.

03

When in doubt, stop. A missing answer is a no, never a quiet yes.

The governance loop

No verdict, no run.

  1. DecidePrudence checks the system's risk tier, its approval and the policy, then answers allow, deny or approval required.
  2. RunAOS, or any framework with a Prudence adapter, asks first and obeys. A missing answer is a no.
  3. ProveDecisions and executions are sealed in Ernest, a hash-chained, tamper-evident record anyone can verify.
  4. OverseePrudence watches posture, drift, incidents and expiring approvals, and they feed the next decision.

Why I built this

I started with evidence and a runtime.
I ended up with a conscience.

CCMLab began in autumn 2025 with Ernest, a layer that keeps verifiable evidence of what AI systems do, and a month later AOS, a Go runtime that runs declared pipelines step by step: YAML pipelines, API calls, agent tasks and human gates. Building AOS taught me how hard running workflows reliably can be—and how many teams and frameworks are already working on that problem.

The questions almost nobody was answering were different. Who allowed this system to run? Under which approval, and is it still valid? What exactly happened, and can anyone prove it a year from now? Those questions are the core of the EU AI Act, and they barely change from one framework to the next.

So the project turned around. Governance and evidence became the centre, and AOS became one runtime that obeys them, alongside any framework with an adapter. Every piece is built to fail closed: if the answer is unclear, nothing runs.

— Carlos, CCMLab

Milestones

One year of work.
One architecture.

  1. Oct 2025Ernest starts: verifiable evidence for AI systems.
  2. Nov 2025AOS starts: a runtime for declared, deterministic pipelines.
  3. Jun 2026Ernest's first public release, v0.1.0-alpha.
  4. Jul 2026AOS v0.3.0, published as a Go module, and Ernest v0.2.0-alpha.
  5. Oct 2026Prudence starts, and the three projects become Chestertown.
  6. Nov 2026Chestertown 0.1.0, all under the Apache-2.0 licence.

The reference implementations

Three roles.
One accountable operation.

PDECIDE · OVERSEE

Prudence

The policy decision point and control plane for AI systems.

  • Registry with EU AI Act risk tiers, including prohibited practices.
  • Two independent sign-offs for high-risk systems; approvals that expire.
  • Incidents with Art. 73 reporting deadlines, posture and drift alerts.
  • Attestations exported as OSCAL and sealed in Ernest.
  • Adapters for LangChain, CrewAI, Google ADK, the Claude Agent SDK and more, and AuthZEN for MCP gateways.
ARUN

AOS

The reference governed runtime and enforcement point: it does the work, and asks before it does.

  • Intent-driven YAML pipelines: linear or DAG.
  • Human gates, callbacks, RBAC and a tamper-evident audit journal.
  • Policy enforcement point for Prudence: deny means nothing runs.
  • A reference, not a requirement: Prudence and Ernest work with your own runtime.
EPROVE

Ernest

The evidence layer: what happened, in an order where any change shows.

  • Hash-chained, tamper-evident record of decisions and executions.
  • Signed submissions; eIDAS-ready timestamps and seals from the trust provider you connect.
  • Kafka ingestion and Merkle batches for volume.
  • Independent verification of any sealed record.
Roles and contracts The reference deployment

The live demo

Five AI systems.
Four risk tiers, and a suspension.

The demo runs AOS and Prudence side by side. Each button in AOS asks to run a different system. One runs straight away; the others stop until someone decides. Approve one in Prudence and the next request in AOS behaves differently, without touching AOS.

Live from January 2027. Each case below has a recording of a real run. Live sessions open from January 2027: message me on LinkedIn and I can walk you through it in a short call, or set up a personal user so you can approve systems yourself.

Demo only. Please do not enter personal, confidential or real-world data. Privacy & demo notice

Minimal risk

Spam check

Allowed from the start: it is already approved, so it simply runs.

Limited risk

Support lookup

Denied until one reviewer approves it in Prudence. Then it runs.

Watch 1:21
High risk

Credit pre-screening

Approval required from two different people: the four-eyes rule.

Watch 1:31
Prohibited

Social scoring

Always denied. Blocked as a prohibited practice, and the policy grants it nothing either.

Watch 0:36
Suspended

Translator

Denied after an incident, until it is reviewed and approved again.

Watch 1:10
Watch all four cases · 4:25

About the model. The demo deliberately uses a deterministic stand-in instead of a real language model. It follows a script and says so in its name, because what is being shown is the governance, not how clever the model is. The demo runs on a small private server, and the code is on GitHub.

Why these names

Why these names

Prudence
A human name with a wink: "prudential" is the word regulators use for keeping risk in check.
Ernest
As in earnest: a record that is honest and serious, and stays that way.
AOS
Agentic Orchestration System. The one that kept its acronym, because it does the work.

A practical premise

Useful AI needs more than a model.
It needs a record of care.

Transcript