Guardrails Belong in the Tool, Not the Prompt

The short answer: when an AI can act on the physical world, its limits have to be enforced outside the model, and its claims have to be checked against the real device. Neither is optional, and neither can live in a prompt.

The setting

Our founder is putting an AI layer on his own home. Home Assistant already ties together a house full of brands — lighting, switches, door locks, televisions, and speakers from different manufacturers — on one control system. The next step is Claude, Anthropic's model, as the interface: you describe an outcome ("I want to watch a movie in the living room") instead of naming devices or remembering what a routine was called.

That's the convenience case, and it is planned, not built. The reliability case showed up during the build, and it's the part we care about.

1. Hard limits live in the tool

The tool the model uses to control the house refuses to unlock a door or open the garage, regardless of how the request is phrased. Locking and closing are allowed; unlocking and opening are done by a person.

That rule is not an instruction the model is asked to follow. A prompt is a request; a tool boundary is a boundary. For anything with physical consequences, only the second one counts.

2. Check the device, not the belief

During one test, the model reported that two speakers were playing. The owner said they weren't. The model queried the speakers directly, got the same answer, and reported it — and then the owner heard them. The model was right, and it was right because it checked the device instead of defending its first answer or deferring to the person.

Earlier the same day, a control tile turned itself off seconds after being pressed while the TV it controlled kept playing. The cause was traced through the system's own event log to a comparison between a number and a piece of text that never matched. It was found by reading the record, not by asking the model to guess.

Confidence is not accuracy, from either side of the conversation.

3. Routines for the predictable, reasoning for the new

Predefined routines are deterministic and fast, so they stay for everyday and safety-critical actions. The model handles what nobody has defined yet. Each mode does the job it is suited to.

4. Scope what the model can see

Home Assistant exposes only selected devices to a conversation agent. The model can't act on what it was never given.

How this maps to Fi

These are the same principles behind Fidenta (Fi), the platform Noevant's products run on. Never assume, verify. An answer comes with its source and how confident we are in it; when models disagree, the disagreement goes to a person instead of being averaged away. The house is a small, physical version of the same discipline.

The full write-up of the build is here: Describe It, Don't Name It.