The scene is the problem

There is a moment, usually about eighteen months after commissioning, that every owner of a serious control system recognises. You are standing in a room. The lights are wrong. And instead of adjusting them, you find yourself performing a small piece of arithmetic: is this worth an email to the integrator?

Almost always, the answer is no. So the lights stay wrong. And a system that cost more than a car quietly becomes a very expensive way of turning things on.

This is not a failure of installation. The integrator did excellent work. It is a failure of the primitive that all of this was built on.

What a scene actually is

A scene is a stored set of output states. Press a button, and eight dimmers go to recorded levels, three blinds go to recorded positions, an amplifier selects a recorded input. KNX implements this beautifully at the device level — scene numbers stored in the actuators themselves, recalled with a single telegram, executing with no controller in the loop at all. Control4, Savant and Crestron wrap richer programming around the same idea.

It is a genuinely good primitive. It is fast, it is deterministic, it survives the processor being unplugged, and it will still work in 2045.

It also has a property nobody likes to say out loud: a scene is a photograph.

It captures one configuration of a house, at one moment, as judged by one person — usually the integrator, on a commissioning afternoon, in a room that is empty of the family who will live in it. That photograph is then replayed, unchanged, for a decade.

The three things a photograph cannot do

It cannot notice that it is wrong. “Evening” at 6pm in January and “Evening” at 6pm in July are different rooms with different light through different windows. The scene does not know which one it is in. It cannot, because it contains no representation of the world — only of the outputs.

It cannot resolve a conflict. Real houses generate contradictions constantly. The vacancy logic wants the west wing off; someone is reading in the annexe. The tariff logic wants the pool pump running now; the household is asleep above it. Scene systems resolve this the way all imperative systems do — whichever instruction arrives last wins. That is not a decision. That is a race.

It cannot be corrected cheaply. Because the logic lives in ETS project files and integrator-only programming environments, the cost of changing behaviour is a site visit. So behaviour effectively freezes at handover, while the household keeps moving. Children are born. Parents move in. A study becomes a bedroom. The house holds its pose.

Why more scenes makes it worse

The standard remedy is more scenes. Morning, Late Morning, Away Short, Away Long, Entertaining, Entertaining Outdoors, Cinema, Cinema Daytime. Every high-end installation eventually accumulates a cast of dozens.

But every scene added is another item in a menu a human now has to select from, and selection is the expensive part. The moment a house has more than about seven recallable states, no one can remember which one they want, and everybody reverts to the two they can find: All On and All Off.

This is the quiet tragedy of the category. Owners spend six figures to arrive at a light switch with extra steps. In the houses we have surveyed, the overwhelming majority of scene recalls come from fewer than four buttons, and a substantial minority of programmed scenes have never been recalled at all — not once since commissioning.

Interface design cannot rescue this. A nicer app is still an app. Voice control is still an instruction. Josh.ai is a genuinely elegant product and it is solving the wrong half of the problem: it makes the command easier to issue. The goal is not to issue commands more gracefully. It is to stop needing to.

The thing the scene is standing in for

Here is what makes this frustrating: the information required to do better is already in the building.

A KNX installation of any seriousness is drowning in signal. Presence detectors, lux sensors, contact states, valve positions, meter readings, setpoint feedback, actuator status objects. In the reference house behind this site, 629 group addresses carry that traffic. The bus knows, at all times, an enormous amount about what is physically happening.

None of it is retained. A presence detector reports occupancy, an actuator switches, and the fact evaporates. Nothing in the system asks how long was this room occupied last Tuesday, or does this room’s occupancy predict the next room’s, or did the owner immediately undo what we just did. The telegram is the whole memory, and it lasts milliseconds.

The scene persists precisely because there is nothing underneath it that accumulates. Automation without memory can only ever be a recording.

What replaces it

Not a better scene. A different shape of system entirely.

Instead of stored output states, you hold a model of the house: rooms as spaces, devices as capabilities, people as occupants with histories. Instead of one script per outcome, dozens of small policies observe that model and each proposes what it believes should happen — the towel rail should begin pre-heating, this zone should hold rather than chase, the dishwasher should wait ninety minutes for the solar peak.

Proposals collide, constantly and by design. That collision is where the intelligence lives. An arbiter resolves them by precedence, so that safety structurally outranks comfort and comfort structurally outranks convenience — not by convention, but because the tiers are enforced in the engine and a lower tier literally cannot win.

And when the house is wrong, you overrule it at the wall, exactly as you always have. The difference is what happens next. In a scene system, your override is noise — an inconvenience the automation will repeat tomorrow. In a system with memory, your override is the single most valuable signal it will receive all week.

The honest objection

A reasonable person reads this and thinks: I do not want my house guessing.

Correct. Guessing is intolerable in a building. The answer is not to make the system more confident, it is to make it accountable — every action carrying the policy that proposed it, the proposals it beat, the evidence it used, and the tier that decided it, written to a log you can read.

That is why we publish our overrule rate, and why we will keep publishing it even in the quarters where it gets worse. A system asking for this much latitude in someone’s home owes them the receipts.

The scene had a good run. It was the right abstraction for an era when the controller had 64 kilobytes and no persistent storage, and it will remain the right fallback layer when everything above it is switched off. But it was always a stand-in for something the technology of the time could not deliver.

The technology can deliver it now.