The boundary between seeing and deciding
Most failures in complex systems are not broken components. They are boundaries that were never properly drawn.
In complex systems, most failures do not come from individual components breaking. They come from boundaries being poorly defined. When responsibilities blur, a system becomes hard to reason about, harder to modify, and fragile under pressure.
The boundary that took the most work to draw properly was the line between perception and decision.
The tempting collapse
It is very natural to merge them. Modern models are powerful enough that you can ask a perception system not only to report what is present but to infer what should happen next. One component, fewer interfaces, less plumbing.
In practice this produces systems that react quickly and reason poorly. The tighter the coupling, the more brittle the behaviour — because the decision inherits every weakness of the perception, silently, with no place in the architecture where you could have caught it.
What the separation buys
Keeping them apart costs an interface and buys three things.
Uncertainty survives the trip. When perception hands over a structured observation rather than a recommendation, it can also hand over how sure it is. A merged component has no natural place to put that, so confidence gets absorbed into the output and disappears. Every downstream safety gate in VECTOR depends on confidence being a first-class value, and that is only possible because the boundary exists.
Failures stay attributable. When a decision is wrong, the separation lets you ask which half was wrong. Did it see incorrectly, or see correctly and choose badly? In a merged component that question has no answer, and a question with no answer is a bug you cannot fix.
Components become replaceable. Perception can be retrained, swapped or degraded without the decision layer noticing, as long as the contract holds. This is what let the system survive its own scope change.
The general rule
A boundary is worth its interface cost when the two sides can fail independently. If perception can be wrong in ways decision-making cannot detect, they are different components, and merging them for convenience means choosing to be unable to tell which one failed.