A machine that learns on the job needs one thing before it needs anything else: a point every action passes through, where the action is allowed, held or stopped. In MA Reflex and MA Duplex, that point is the gate.
Three answers, on every action
- Commit. The action is confident and inside every rule, so it runs.
- Defer. The runtime is not sure, so the decision goes to the policy, the model or a person.
- Refuse. The action is unsafe or outside a limit, so it is not carried out.
Confidence earns autonomy. Uncertainty asks for help. Anything outside the rules never happens. Every one of those answers is written to the record, so an operator can see what the machine did and why.
Fail closed
When the gate cannot reach a confident decision, the safe answer is no action, not a guess. A gate that fails open is a report, not a control. MA Reflex fails closed: if anything in the decision is missing or out of bounds, the action is held.
Limits set at runtime
Keep-out zones, limits and policies are part of the decision itself, and they can be added or released while the machine works. Close a zone and the next action already respects it. Nothing outside them is carried out.
Alongside existing safety systems
The gate does not replace a hardware e-stop or a safety controller. It runs alongside them. The hardware keeps people safe from the machine. The gate keeps the machine's own decisions inside the rules, on every action, and keeps the record that lets a team sign off on autonomy.



