The reference deployment

The architecture,
running.

The architecture is roles and contracts. This is what it looks like running: Prudence, AOS and Ernest deployed together, with one sign-in, one way of working, and every run traceable from the request to its evidence. Each one still installs and runs on its own.

DECIDE · OVERSEE · PRUDENCE

Approves and oversees AI systems.

RUN · AOS

Runs governed AI pipelines.

PROVE · ERNEST

Tamper-evident audit evidence.

What the deployment adds

Three consoles.
One way of working.

01

One sign-in

A shared identity provider signs you in to all three. Your name and roles follow you from one app to the next.

02

One switcher

The same bar sits at the top of every console. It shows where you are, and says in a few words what each app does.

03

One design

A shared design system: the same buttons, cards, alerts and verdict badges, in light and dark, on a desktop or a phone.

04

Linked records

A run links to the AI system and decision that allowed it. Its evidence links back to the exact run. No copying ids between tabs.

05

Upgraded one at a time

Each app is deployed on its own schedule. Updating one never forces the others to stop.

Follow one request through the deployment

Follow one request through the deployment

  1. AskIn AOS, press Support lookup. Before any step runs, AOS asks Prudence whether this AI system may run.
  2. DecidePrudence answers allow, deny or approval required. A blocked result names the system and opens it in Prudence.
  3. ApproveA reviewer approves the system in Prudence. Nothing changes in AOS: the next request is simply allowed.
  4. ProveThe run shows Governed by Prudence, with the decision. Evidence in Ernest opens its sealed record.
  5. ReturnIn Ernest, each event from AOS carries Run in AOS, back to that exact run. The loop is closed.

Open the deployment

Start anywhere.
The switcher takes you on.

PDECIDE · OVERSEE

Prudence

Approves and oversees AI systems: the registry, approvals, incidents and posture.

ARUN

AOS

Runs governed AI pipelines, and asks Prudence before every governed run.

EPROVE

Ernest

Tamper-evident audit evidence: decisions and executions, sealed and verifiable.

Use one, or all three

Use one implementation, or all three

Prudence
Can govern other runtimes too: adapters for agent frameworks, from LangChain and CrewAI to the Claude Agent SDK, ask it first, just like AOS. MCP gateways can ask it about each tool call.
Ernest
Accepts evidence from other systems as well, over Kafka, CloudEvents or MLOps connectors.
AOS
Stays small on purpose: the reference runtime that shows how a runtime obeys the contracts. Nothing here requires it.

eIDAS

Built for EU trust law.
Without borrowed authority.

The AI Act says keep records; eIDAS says how records earn legal trust.

01

Qualified timestamps

Ernest timestamps its evidence with RFC 3161 time-stamping authorities. Use a qualified trust service provider, and the timestamps are qualified.

02

Qualified seals, ready to connect

Submissions can carry qualified electronic seals created by a provider's remote-signing service (CSC) and checked against the EU Trusted List.

03

No borrowed authority

Chestertown is not a trust service provider. Qualified status comes from the provider you connect; without one, the evidence is cryptographically verifiable but not legally qualified.

The loop in one line

Decide, run, prove, oversee.
In that order, every time.