Skip to content
Autonomy you can answer for

The runtime gate between AI intent and real-world execution.

AAXO is a cognitively inspired governed runtime between an AI system's reasoning and its real-world action — binding observation, decision, policy, execution and evidence into one controlled loop. Hand AI real work without handing over command, and answer for the result afterwards.

aaxo · runtime gate
AI intentagent.intent()

Policy check

evaluate · verify · bound

Approve
Degrade
Escalate
Block
Action + evidenceevidence.write()
GateNothing executes unchecked
TraceEvery decision on record
VerifyOutcomes confirmed, not assumed
BoundLimits set by people, enforced at runtime
Why now

AI agents are moving from chat to action.

The risk has moved from wrong answers to uncontrolled execution. AAXO controls the execution boundary.

When an agent takes a wrong action today, there is often no record of why it acted, no one accountable for the decision, and no way to demonstrate to a board, auditor or regulator what happened.

ChatAction

AI agents are no longer just generating text. They call tools, edit code, touch infrastructure and operate workflows.

Wrong answersUncontrolled execution

The risk has moved from bad output to real-world actions that touch production systems, customers and physical equipment — taken without control, evidence or recourse.

Model trustArchitectural control

Control belongs at the execution boundary — enforced by the runtime, not promised by the model.

Model-independent

Your rules should not depend on someone else's model.

The same rules, enforced the same way, every time — regardless of which model is reasoning. That makes governed autonomy viable on-prem, air-gapped and at the edge: autonomy that stays under your command, on your infrastructure.

The question is shifting from what AI can do to what your organization can answer for.

The problem

AI can now act on your behalf. The consequences are still yours.

Today's agents can plan, write code, analyze data and operate tools. As autonomy increases, they often lack robust control, verifiability and governance when they move from suggestion to action.

Capability without control is unacceptable.

In serious environments — enterprise, industry, defence, robotics — the questions a responsible operator must answer remain unanswered by the model alone.

If it acts in your name, you should be able to answer these
  • 01What did the AI system observe?
  • 02Which decision did it make, and why?
  • 03Which action did it attempt to execute?
  • 04Which rules and safety boundaries were checked?
  • 05Was the action approved, blocked or degraded?
  • 06Is there reviewable evidence after the fact?
AAXO is built so each one has an evidenced answer.
Principles

AAXO exists so organizations can adopt autonomy without surrendering command.

People set the policy. The runtime enforces it. The evidence remains on record.

Humans hold final authority

Actions your policy marks consequential escalate to a named owner.

Every action is answerable

If it cannot be explained afterwards, it should not execute.

Control is enforced, not promised

The runtime decides, not the model's good intentions.

Sovereign by design

Built to run where you run: on-prem, air-gapped, at the edge.

The safety belt, braking system and flight recorder for operational AI.

Architecture

A cognitive control architecture for AI that acts.

AAXO does not simply approve or block tool calls. It structures the full operational loop: what the system observed, how it reasoned, what it intended to do, which policy applied, what action was released, and what evidence records the outcome.

Observe
Reason
Decide
Govern
Act
Verify
Audit
Learn

A closed loop: every stage is observable and governed, not a black box. The execution gate is one control point within it.

The execution gate

Every AI action passes through a governed cognitive loop.

The execution gate is one control point in that loop — the boundary where each intent is evaluated against policy and resolved to a single accountable outcome, with a verifiable record behind it.

AI intentDecision from the modelagent.intent()
aaxo · runtime gate

Policy check

evaluate · verify · bound

Approve
Degrade
Escalate
Block
Action + evidenceExecuted within boundsevidence.write()
Approve

Intent matches policy — action passes through unchanged.

Degrade

Action is narrowed to a safe, bounded version before release.

Escalate

Decision is routed to a human owner for explicit approval.

Block

Unsafe or out-of-policy action is stopped before it executes.

Whatever the outcome, a human owner can always see why — and your policy decides which actions require one's explicit approval first.

The AAXO layer

A governed runtime between the agent and the real world.

Delegation should not mean abdication. The agent proposes; your policy decides; the evidence remains.

AAXO is not a dashboard that reports what an AI did afterwards. It is an active runtime that can validate, constrain, degrade or stop actions before and during execution. It can prevent an action before it happens — not just explain it afterwards.

upstream

Model / Agent

Plans, reasons, calls tools and proposes actions.

  • LLM reasoning
  • Tool calls
  • Multi-step plans
AAXO runtime

Governed execution layer. Every proposed action passes through policy, gating and evidence before it reaches a real system.

Policy engine
Execution gate
Audit trail
Evidence chain
Human oversight
Safe degradation
downstream

Tools / Systems

Databases, APIs, infrastructure, machines and operations.

  • Production systems
  • Physical actuation
  • Critical ops
Core capabilities

The controls that make autonomy accountable.

Each one answers a question of trust: what may run, who decides, what is on record, and what happens when limits are reached.

Decide

Policy-gated execution

No silent execution outside defined rules — every action is evaluated against explicit policy before it runs.

Decide

Agent action boundaries

People decide the limits — clear rules for what an agent may do, what needs approval, and what must be blocked.

Decide

Runtime verification

Trust is checked while it matters — state and constraints validated during execution, not only after the fact.

Prove

Audit trails

Answerable after the fact — every meaningful decision and action recorded with the context it was made in.

Prove

Evidence-linked decisions

Review does not depend on taking our word — decisions are bound to verifiable evidence external parties can review.

Prove

Operational traceability

The full story is reconstructable — the path from observation to action, across the agent lifecycle.

Command

Human oversight

People keep the final word — defined points to inspect, approve or halt.

Command

Safe degradation

Failing safely is a design stance — outside bounds, the system steps down to a safer mode instead of failing open.

Use cases

Built for environments where actions have consequences.

Wherever an AI action touches the real world, a person is responsible for it. AAXO is built for those people.

  1. Software repair & code execution

    01

    Automated changes to code and infrastructure gated by verification and evidence.

  2. Enterprise AI agents

    02

    Agents acting on internal systems with policy, approval and a full audit record.

  3. AI governance & compliance

    03

    Verifiable evidence that systems followed their rules — auditable after the fact.

  4. Industrial automation

    04

    Operational AI inside control loops with defined limits and human oversight.

  5. Robotics & physical systems

    05

    Bounded actuation where unsafe actions degrade or stop before reaching hardware.

  6. Defence & critical operations

    06

    Traceable, controllable autonomy for environments where consequences are real.

Evidence

Technical validation, without overclaiming.

AAXO is being validated through evidence-based evaluation and controlled testing. We report what the runtime demonstrably does — and where the boundaries currently are.

We hold our own claims to the same standard the runtime holds AI actions: gated by evidence. Evidence, not self-report.

  • Early runtime validationControl layer exercised in controlled execution scenarios.
  • Model-in-loop testingAgents evaluated with governance active in the decision path.
  • Controlled benchmark environmentsReproducible setups for measuring gated vs. ungated behaviour.
  • Evidence-based evaluationOutcomes assessed against recorded evidence, not model self-report.
  • Audit-chain verificationDecision and action records validated for integrity end to end.
Current validation status

Early technical validation is in place. Enterprise pilots are under preparation. We are deliberately precise about what is proven today versus what is in progress.

Proven

Runtime gating in controlled scenarios

Measured

Attack vs. benign intent discrimination, over-refusal boundaries and fallback behaviour — detailed results available under review

In progress

Enterprise pilots under preparation

Next

Production deployments with design partners

Not claimed: production safety certification · unbounded autonomy · model capability guarantees

From AI output to governed action

The safety belt, braking system and flight recorder for operational AI.

Not to slow AI down — to make it something your organization can stand behind. For enterprises, partners and investors building operational AI they intend to answer for.