Autonomous Orchestration & Reasoning Architecture

Governance
for AI agents
that actually works.

A:O:R:A helps companies understand and control what their AI systems are actually doing.

Open-source. Self-hosted. Every AI decision passes through your rules — not the model's instincts. Built for teams in regulated environments.

View on GitHub Why this matters ↓
Core principle  —  Policy decides. Not the model.

AI is moving fast.
Accountability is catching up.

Companies are deploying AI that books appointments, answers medical questions, makes financial recommendations — and takes the blame when it gets things wrong. Courts, regulators, and insurers all agree: if your AI says it, you said it.

Urteil
Chatbot hallucinates — operator is liable.

OLG Hamm ruled 12 May 2026: a chatbot is not an independent third party. Its statements are the company's statements — regardless of whether the AI hallucinated or processed correct data incorrectly. The operator bears full responsibility.

OLG Hamm · Az. 4 UKl 3/25 · 12.05.2026 · Revision BGH zugelassen
2026
EU AI Act — mandatory auditability

High-risk AI systems now face mandatory transparency, logging, and human oversight requirements. Most existing AI stacks have no answer for this.

EU AI Act · In force since Aug 2024 · Full application 2026
0
Existing tools govern nothing

LangChain, AutoGPT, CrewAI — powerful tools, but none enforce policy. They route. They chain. They do not govern. A:O:R:A fills exactly that gap.

Market analysis · 2026

01 — What is A:O:R:A

The governance layer
your AI stack is missing.

Think of A:O:R:A as a control room: every AI request goes in, every response comes out — and your rules decide what is allowed in between.

Every request passes through a policy enforcement layer before it reaches a model. If a rule is violated, A:O:R:A blocks. It does not guess. It does not pass through.

It is model-agnostic: works with Claude, GPT-4o, Gemini. Switch providers without rewriting your policies.

Policies are defined in plain YAML — readable by anyone. Every decision is written to an immutable audit log. Nothing deleted. Nothing edited.

Designed with EU AI Act and NIST AI RMF principles in mind — built for regulated environments.

Claude GPT-4o Gemini + more
// policy_example.yaml
version: "1.0"
mode: controlled
 
# Block on violation
on_violation: block
audit: append_only
 
intent_rules:
  - pattern: research
    allowed: true
  - pattern: data_exfil
    allowed: false
    risk: CRITICAL
 
human_in_loop:
  threshold: HIGH
  require_approval: true

Six foundations.
None of them optional.

01
Policy decides. Not the model.

Your rules run first. Always.

No AI model determines what is allowed. A:O:R:A enforces your YAML policies before and after every model call.

02
Fail-closed by default.

When in doubt, block.

If a decision cannot be made with certainty, A:O:R:A blocks. It never defaults to open.

03
Immutable audit trail.

Every decision recorded. Nothing disappears.

Every action is logged — append-only. Compliance reviewers see exactly what happened, when, and why.

04
Human-in-the-Loop.

High-risk decisions need a human yes.

Where the stakes are high, A:O:R:A pauses and requires human approval before proceeding.

05
Model-agnostic.

Change AI provider. Keep your policies.

Claude, GPT-4o, Gemini — the governance layer stays the same regardless of which model runs underneath.

06
Fully self-hosted.

Your data stays with you. Full stop.

No telemetry. No cloud dependency. No vendor lock-in.


03 — Capabilities

What A:O:R:A does.

Governed Pipeline

Every AI call follows a fixed, auditable path.

Input Gate → Policy Evaluation → Model → Output Verification. No shortcut. No bypass. No exception.

Intent Classification & Routing

A:O:R:A understands the purpose of a request before acting.

Research, agentic tasks, simple queries — each follows a different governed path with matching risk controls.

Safety & Watchdog Runtime (SWR)

Continuous monitoring. Anomalies trigger escalation.

Three configurable safety profiles (light / controlled / full) monitor agent behaviour in real time.

Persona Leak Filter

A:O:R:A identifies itself correctly. Jailbreaks rejected.

The underlying AI provider's identity never surfaces to end users. Manipulation attempts blocked.

Council Review Protocol (CRP)

Contested decisions go through structured review.

Multi-layer verdict system with escalation logic for high-stakes situations.

Deep Research Mode

Multi-step research — governed end-to-end.

Research loops with evidence aggregation, deduplication, and confidence labelling — under policy control.

Local-First Knowledge Evidence

A:O:R:A checks local evidence before trusting generated answers.

LFKE makes internal documents, verified knowledge sources, and evidence status visible — including confidence, source usage, and block reasons when evidence is missing or weak.

Decision Transparency

Not just answering. Showing why.

A:O:R:A shows whether a decision was allowed, blocked, or unavailable — with reason and a human-readable explanation. Don’t trust us. Verify us.


04 — Demo

See it in action.

[ COMING SOON — DEMO VIDEO ]

A walkthrough showing A:O:R:A blocking a real policy violation in real time will appear here.


Governance without
killing performance.

Every request is governed. The overhead is real but controlled — and we publish the numbers honestly.

690+Passing Tests
~11sAvg. Latency
3Live AI Providers
AGPL-3.0License
[ COMING SOON — SCIENTIFIC BENCHMARK REPORT ]

Independent benchmarking in collaboration with academic partners. Results published here.


06 — Architecture

Five steps. Every time.
No exceptions.

Every AI request follows the same fixed, auditable path through A:O:R:A.

01
Input Gate

Request arrives and is cleaned.

Sanitised and classified for intent and risk before anything else happens.

02
Policy Evaluation

Your rules run. ALLOW, BLOCK, or ESCALATE.

YAML policies evaluated. A CRITICAL verdict is a hard stop — no override.

03
Human-in-Loop?

High risk? A human must say yes.

High-risk decisions wait for human approval before the model is called.

04
Model Call

Allowed requests reach the AI provider.

Only policy-cleared requests reach Claude, GPT-4o, or Gemini.

05
Output Verification

Response checked before delivery.

Model response validated against policy. Audit log written — immutably.


07 — Use Cases

Who builds with A:O:R:A.

[ TO BE FILLED — USE CASE DESCRIPTIONS ]

Planned: Enterprise AI teams, regulated industries (finance, healthcare, legal), compliance officers, and researchers who need auditable AI workflows.


08 — Why this was built

AI without control
is not progress.

“AI systems are being deployed into real decisions — financial, legal, operational — with no reliable way to verify what they actually did, or why. That is not a technical problem. It is a trust problem. A:O:R:A is my answer to it.”

A:O:R:A was built by a solo founder who believes the current wave of AI deployment is moving faster than the governance infrastructure that needs to support it. Designed from day one around one principle: policy decides, not the model.

— Dirk Müller  ·  Founder & Creator, A:O:R:A  ·  Berlin

The world is moving
in this direction.

Courts, regulators, and industry bodies are independently arriving at the same conclusion: AI governance is not optional. It is a legal and operational necessity.

OLG HAMM · GERMANY · MAY 2026
If your chatbot says it — you said it. Even when it hallucinates.

The court ruled that a company's AI chatbot is not a "third party". Every output is attributed to the operator, regardless of whether the AI hallucinated or processed data incorrectly.

Az. 4 UKl 3/25 · Revision BGH zugelassen
Verbraucherzentrale NRW v. Aesthetify GmbH · 12.05.2026
EU AI ACT · 2024–2026
High-risk AI must be transparent, logged, and under human oversight.

The EU AI Act requires high-risk AI to maintain logs, support auditability, and ensure human oversight. A:O:R:A is designed with these principles as a structural foundation.

In Force · Full Application 2026
Regulation (EU) 2024/1689
NIST AI RMF · USA
A framework for managing AI risk across the full lifecycle.

The NIST AI Risk Management Framework establishes governance, mapping, measurement, and management as core pillars. A:O:R:A operationalises these at the infrastructure level.

National Institute of Standards and Technology · AI RMF 1.0
COMING SOON
Press coverage and academic partnerships.

Coverage, research citations, and university benchmark collaborations will be listed here as they are confirmed.

[ TO BE FILLED ]

10 — Open Source

Built in the open.
Auditable by anyone.

A:O:R:A on GitHub

The full governance layer — policy engine, audit system, SWR, CRP, intent router — is publicly available. Fork it. Audit it. Deploy it. No black boxes.

AGPL-3.0
View Repository →

11 — Business Model

How A:O:R:A is sustained.

[ TO BE FILLED — BUSINESS MODEL & COMMERCIAL OPTIONS ]

Open source core (AGPL-3.0). Details on commercial support and enterprise licensing will be published here.


Get Started

Your AI stack deserves
a governance layer.

A:O:R:A is free, open source, and ready to deploy.
Policy decides. Not the model.

Start on GitHub Contact the Founder