Why the planner stays classical
The most consequential architectural decision in the quantum work is about where a component is not allowed to go.
In both QANTIS papers the quantum processor handles a bounded belief-update primitive, and the planner that consumes its output is classical. So is the action authority. People sometimes read this as a concession, a placeholder until hardware improves. It is not. It is the decision I would defend hardest.
Three reasons, in increasing order of importance
Availability. Quantum backends queue. They calibrate. They go offline. A decision loop that stalls when a backend is unavailable is not a decision loop, it is a research demo with a scheduling dependency. Keeping the planner classical means the system degrades to a path that was always there rather than stopping.
Reasoning. I can reason about the failure modes of a classical planner. I can bound them, test them, and explain them to someone who has to sign off on the system. The equivalent story for a quantum planner does not exist yet, and pretending otherwise in a domain where the output authorises physical action would be indefensible.
Authority. This is the one that matters. Action authority is a property of the system that someone has to be accountable for. Placing it inside a component whose behaviour is probabilistic in an unfamiliar way, executed remotely, on hardware whose calibration state changes between runs, moves accountability somewhere it cannot sit.
What a bounded primitive buys you
Once the quantum part is bounded, several things become tractable at once.
You can compare it against an exact classical reference, because the input and output are both well-defined distributions. You can test whether it survives repeated use, which is what the sequential study did across 8, 12, 20 and 32 steps. You can put an evidence gate in front of it. You can remove it without redesigning the system.
That last property is the one people underrate. An architecture where the interesting component can be taken out is an architecture you can be honest about. If the quantum path never clears G3, NODERIQ still has a classical AI core that does useful work. The programme does not depend on the research succeeding.
The version of this I disagree with
There is a school of thought that says bounding the quantum component this tightly gives away the upside. If the quantum layer only ever handles one primitive inside a classical loop, the argument goes, you have capped what quantum can contribute before you have found out what it could do.
I take that seriously and I still think it is wrong, for a practical reason. The unbounded version is not testable. You cannot compare it against a classical reference because there is no matching classical formulation. You cannot attribute a result to the quantum layer because the layer is entangled with everything else in the system. You end up with a system that either works or does not, and no way to say why.
A capped contribution you can measure is worth more than an uncapped contribution you cannot.
Where this shows up in NODERIQ
The same reasoning produced the separation between recommendation and authority at the programme level. A system may recommend. A person authorises. The quantum question is a special case of a more general commitment: components that carry uncertainty feed decisions, and decisions that carry consequence route to someone accountable.
Getting that boundary right in the small case, a single belief-update primitive, is how you learn where to put it in the large one.