SENSING
Crowd density is not engagement. The gap between a signal and a meaning, and where show logic should stop trusting it.
A depth camera reports that fourteen bodies are in a zone. That is a measurement and it is probably accurate. The temptation, immediately and always, is to treat it as information about how the show is going — that fourteen people are enjoying it, that the room is working, that the content should adapt. None of that is in the signal. The signal says fourteen bodies.
The distance between a measurement and a meaning is where most adaptive experiences fail, and they fail invisibly. A system that responds confidently to a misreading looks exactly like a system that is working, right up until someone stands still for a reason nobody modelled and the room decides they are captivated.
The failure is worse than a simple wrong answer because it is confident and it is silent. A misread does not raise an alarm. The room simply does something that made sense to the model and no sense to the person standing in it, and because the room is behaving smoothly, everyone assumes it is behaving correctly. Debugging that after opening means arguing about intent rather than about a fault.
The discipline is to write down, before any code, what each signal can and cannot support. Presence is well measured. Dwell time is well measured. Attention is inferred and should be labelled as inference wherever it appears downstream. Emotion, from anything short of a controlled setup, is a guess dressed as a number, and putting it into show logic means the show is now guessing too.
That does not mean sensing is not worth doing. It means the show should degrade sensibly when the reading is wrong, which is a design decision rather than an engineering one. Adaptive behaviour that makes a good experience slightly better is a fine use of an uncertain signal. Adaptive behaviour that makes a good experience unrecognisable when the signal is wrong is not, however impressive the demonstration.
It helps to hold two separate ideas apart in the specification: what the sensor measures, and what the show does about it. These get collapsed constantly — a feature described as "the room responds to engagement" has already assumed engagement is a thing being measured, when what is being measured is bodies and time. Writing the two lines separately forces the inference into the open where it can be argued with.
The question to ask of any adaptive feature is simple and unpopular: if this sensor were producing noise, would anyone in the room be able to tell? If the answer is no, the feature is decoration. If the answer is yes, you now know exactly what has to be right.
There is also a privacy dimension that is easier to settle at design time than at review time. Cameras in a public space invite a reasonable question about what is stored, and the answer is much stronger when the system was built to derive a count and discard the frame than when it was built to record and analyse later. Deciding that early costs nothing. Retrofitting it costs a rebuild.