A critical step
towards spatial
intelligence.

Life does not arrive
as a list of commands.

“I want to read for a while.” The intention arrives naturally. The work of deciding which light to adjust, how far to close the blinds and whether the room feels comfortable does not. Today, much of that translation still belongs to the person using the system.

Traditional scenes capture a useful state, but real life keeps moving. The same activity happens at different times, with different people nearby, under different temperatures and levels of daylight. A fixed device snapshot cannot, on its own, express all of that context.

Our opportunity is to carry more of the translation inside the system. The question becomes: what environment would support this behaviour, here, now, for this person?

Make the environment
a first-class idea.

Between human intention and device execution, we propose an environment layer. It describes the qualities we want to create: light, glare, temperature, privacy, safety and atmosphere. Devices become the means by which those qualities are achieved.

This separation matters. A preference for reading light should remain meaningful when a lamp is replaced. A comfortable evening should be understandable without requiring the person to name every circuit. Language interprets the request; spatial context and known capabilities shape an executable plan.

Scenes still have a place. They provide a visible, predictable representation of an intended state—something a person can inspect, adjust and return to. Over time, they can become outcomes of learning as well as tools for deterministic delivery.

The first breakthrough
may be almost invisible.

The immediate step towards this future is to capture the right evidence. A person saving a state they like gives us something unusually valuable: an environment they have judged for themselves. An execution record tells us when and where it was used. A subsequent adjustment tells us where the system’s assumption diverged from their preference.

These are different observations, and their structure matters more than the novelty of the interface. A snapshot should preserve semantic states—device, capability, target value and unit—alongside its creation context. A raw protocol command records an instruction; a structured capability state preserves something that can later be interpreted.

A useful execution record carries the surrounding context, before-and-after state and relevant corrections. An illustrative observation window might follow execution for thirty seconds, but the window and the meaning of a change require validation. A correction is evidence to interpret, not automatic proof of dissatisfaction.

“Cozy” and “that worked”
are different data.

A label attached to a saved snapshot defines a person’s vocabulary. It helps the system learn what that household means by “cozy”, “reading” or “dinner”. An evaluation attached to a particular execution answers another question: did this outcome fit this context?

Merging the two loses the distinction between a definition and an experience. Light edits carry meaning too: lowering a saved brightness value should leave a correction history, rather than silently replacing the only evidence of what changed.

The interaction should remain light. Saving, naming or adjusting a state should help the person immediately. Feedback should be optional. Inspectable plans, understandable room and device names, and control over what happens are the foundation for trust.

A home should have
a useful beginning.

Personalisation cannot begin by asking people to train a blank system. Design supplies rooms, zones, devices and relationships. Delivery supplies tested baseline scenes. Simple preference choices can establish an initial direction. Later corrections can refine that baseline.

Our behaviour-centred proposal treats those stages as one continuous preparation path. The aim is a useful first day and a better-informed first week—not a promise that an unproven learning system will immediately understand every household.

Build the conditions
for intelligence.

The hard work lies in making meaning travel: from a floor plan to a room, from a room to a capability, from a capability to an experience, and from that experience to evidence we can learn from. It spans design, delivery, product and engineering.

Success in this first phase should be judged by coverage, structural correctness, contextual completeness and the quality of captured corrections. Activity counts alone do not tell us whether the foundation is sound.

The long-term ambition is an environment that becomes easier to live in as it learns what matters. The practical work begins with a smaller commitment: preserve the meaning of what people already tell us, through the states they save and the changes they make.