01 · INTENT
The agent proposes an action
An AI system produces an intent: subject, operation, parameters, target. Nothing has fired yet.
rover-7 · THRUST_ADJUST · Δv 9.5 · G1
Solutions · Autonomous systems
Give autonomous systems a clear limit on what they may do.
A machine can propose a command that exceeds its permission. The recorded rover decision below shows the kernel refusing one such request. That same check-before-action pattern can be applied to other connected machines with their own rules and integrations.
Watch one decision
A real refusal, stage by stage.
This is not a mock-up. Every strip below is read from a decision the kernel actually made and signed: a burn command of delta-v 9.5 against grant G1's bound of 0.5.
02 · BLUEPRINT
The intent is canonicalized
The request is reduced to a deterministic form, so the same intent always evaluates the same way.
intent hash 6d3b2ebce1baa7891fdc4835cb998033…
03 · POLICY
The policy bundle evaluates it
Every rule runs before anything executes. This trace is from the recorded decision, rule by rule.
G-struct ALLOW · G-epoch ALLOW · G-temporal ALLOW · G-catalog REFUSE · G-grant ALLOW · G-bounds REFUSE
04 · AUTHORITY
The grant is checked
Authority is a held grant with bounds. This intent claims grant G1, which authorizes delta-v up to 0.5.
grant G1 · bound Δv ≤ 0.5 · requested 9.5
05 · DECISION
REFUSERefused, before execution
Out of bounds, the action never reaches the thruster. This one was nineteen times the authorized maximum.
PARAMETERS_INVALID · EFFECT_BOUND_EXCEEDED
06 · RECEIPT
A signed receipt, either way
The decision, the intent hash, and the rule trace are signed together. Anyone can verify it without trusting ExecLayer.
rcpt-G1-msi6fhqi · ED25519 · 2c9f43e94821c35f5f5a650b…
The receipt this walkthrough prints is the recorded artifact itself. Watch the kernel refuse a live intent.
How it fits into your organization
Who owns it, and how it connects.
You keep your AI and existing systems. We start with one workflow, configure checks around your approved rules, and connect them where the AI tries to act. Authorized requests can proceed. Other requests stop, with a record explaining the decision.
- 01 · Set the rules
Business and control owners
Your Robotics Program Lead or Operations Director defines permitted actions and operating limits, supported by safety engineers.
- 02 · Connect the workflow
Integration specialists
Your robotics and controls engineers connect the command workflow.
- 03 · Check before action
The connected execution path
Before a command reaches the connected machine controller. The integration must prevent the agent from going around the check.
What does a first implementation involve?
We scope one workflow, translate supported rules into a versioned, validated, signed policy package, and connect an adapter that supplies trusted facts such as identity, recipient, and approval status. We then test allowed requests, blocked requests, and attempts to bypass the check before enabling execution.
One maintained kernel can support different policy packages and system adapters. A JSON file alone is not a complete deployment. We confirm rule support and integration requirements for your systems during scoping; additional implementation may be needed.
These are typical ownership patterns; responsibilities vary by organization.
Why this check matters
A proposed action exceeds the permission it holds.
The recorded request asks for a delta-v of 9.5 while its grant permits no more than 0.5. The kernel refuses it. Applying this pattern to a robot or another autonomous system requires that system’s operating limits, trusted inputs, and a command path that cannot bypass the check.
ENV · AUTONOMOUS SYSTEMS
Related products · distinct workflows
Explore the product for the workflow.
The recorded rover decision demonstrates the permission-checking mechanism. QueueFlow Brain applies authority boundaries to machine orchestration; SovereignClaw applies them to AI agent actions. Each has its own connected workflow.
QueueFlow Brain
What it is
A robot-orchestration kernel that decides what coordinated machines are permitted to do.
The problem it solves
Orchestration multiplies authority: one instruction becomes many physical actions across many machines. QueueFlow Brain evaluates each against its operating limits before dispatch, so an unauthorized action is refused mechanically rather than caught by a person watching a screen.
SovereignClaw
What it is
A deterministic AI agent that mechanically refuses unauthorized execution.
The problem it solves
An agent holding real credentials will eventually be steered somewhere it should not go, whether by a bad input or a plausible chain of reasoning. SovereignClaw treats permission as something enforced at the execution boundary rather than something the model is asked to respect.
The evidence it leaves
A decision is only settled if someone else can check it.
The walkthrough above reads a captured kernel receipt. It shows what was requested, which rules were evaluated, and why the request was refused. Open the live rover demonstration to make a new request and inspect its decision, or the labeled captured fallback if the live service is unavailable.