Systems that cannot fail, not usually succeed.
Four questions decide it. If you can answer all four about what you are building, we can almost certainly help.
Does something change over time?A quantity that evolves — a glucose level, a delivered dose, a tool position, a charge. Not a single input scored once.
Is there a number it must never cross?A maximum rate, a target band, a geometric boundary. Not a threshold the system crosses as its ordinary function.
Does an output reach a patient without a person in between?An actuator, a pump, a stimulator, a commanded setting. Where a clinician reads and decides, the risk lives in their judgement.
Is there a model, or could there be one?Controller equations and a plant abstraction you would put your name to. You do not need it written yet.
Where that shows up.
Closed-loop control
A quantity is measured, a decision is made, and the decision changes the quantity. Nothing outside the loop is watching, and the loop runs unattended.
Learned decision functions
An AI model whose behavior was fitted rather than written, deployed where its output is acted on rather than merely reviewed.
Autonomous monitoring
Systems whose job is to notice a condition, where a missed one is the failure that matters and testing can only ever show it did not happen yet.
A strong submission carries three kinds of evidence.
Process
Your QMS and validation consultantsThe right procedures were followed.
Statistical
Your CROYour AI model performed well in this sample.
Mathematical
SKEPISA proof that the system model containing your AI cannot exit the declared bounds.
All three belong in a submission. We do one of them.