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
Autonomy is not the risk. Unbounded authority is.
Autonomous systems plan and act faster than any human review loop can follow. ExecLayer bounds what they are permitted to do at the moment they try to act, without slowing down what they are able to do.
What goes wrong
Review that cannot run at machine speed
An autonomous system decides and acts in milliseconds. Any safety story that depends on a person looking at something has already conceded it does not work at production tempo. The boundary has to be automatic, and it has to be in the execution path rather than beside it.
ENV · AUTONOMOUS SYSTEMS
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.
What we build for it
The products built for this work.
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.
At machine tempo the record is the only account of what happened that anyone will ever have. It has to be produced by the decision itself, not written about the decision afterwards.