← All work

VeroField study

What is the automation system in my house actually doing, and can software act on it without guessing?

Active Ongoing engineering study.

A safety-bounded reverse-engineering study of a professionally installed home automation system.

The problem

The house I live in runs a professionally installed home automation system: lighting, shades, climate, audio and video, and scenes. Installers set it up. I live with it every day, but nothing written said what it actually does, and it was never built for software to drive.

AI agents are starting to act on real things, and a home is the hard case because people live in it. The danger is not only a bad change but a wrong report. A command the system accepted is a receipt, not proof that the light moved, and an agent that trusts the receipt retries or tells a person something false.

What I built

  • Read-only instrumentation first, on a known network with the homeowner's knowledge. It subscribes to the system's live state feed and records what changes. No write path existed until the reading was understood.
  • A guarded write path for four allow-listed lights. Each request is checked, the light is read first, the change is sent once, and the light is read back until it settles.
  • Exactly one of four outcomes for every attempt: confirmed, unconfirmed, failed or refused, each with its reason.
  • A fake house calibrated from 37 recorded real writes, so every test and the demo run without touching the real one.
  • A Swift package with no dependencies: the VeroCore runtime and a vero command-line tool, tested with XCTest.

Built with Claude Code agents under my direction, in the house I live in. The written record labels every finding as measured on the house, published by the vendor or inferred, and keeps four of its own wrong conclusions on file.

Decisions, and what broke

  1. Accepted is not done

    The system answers a write before the light has moved, so Vero never treats that answer as success. It reads the light back, and when the reading can't prove the result it says unconfirmed, which tells the caller not to retry blindly.

  2. Call it failed only with proof

    A probe found that an unrecognized write error was reported as failed with zero writes, while the light had in fact moved to the requested level. Now failed requires an error that proves the write never got through; any other error is read back and never retried. Five new tests failed on the old code before the fix, then passed.

  3. Refuse when the map is out of date

    Before any write, Vero reads the light's binding from the system and compares it with its saved model. If the installation was reconfigured since, the request is refused before anything is sent.

  4. Tests never touch the house

    Vero reaches a house only through one three-call seam: read a state, set a level, read a binding. A fake adapter fills that seam in every test and in the demo, which prints each of the four outcomes once.

  5. The live feed speeds up checking, not sending

    Pushed updates tell Vero sooner that a light moved, with polling as the fallback. Plain HTTP stays the only way a change is sent, because writes over the socket were never tested.

Evidence

Select a claim

Corrected

1.5seconds · assumed
0.15–0.41seconds · measured
Twelve timed writes
The simulated house first assumed a light takes 1.5 seconds to settle at a new level. Twelve timed writes on the real house showed it within 0.15 to 0.41 seconds.

Proof

24 live writes, 0 failed
On 4 October 2026, on the four allow-listed lights, with the homeowner home and watching. Every light read back at its starting level afterwards.
1,618 of 1,632
Apparent changes in one full sweep of the state feed that were only the same value restated in another JSON type. Vero compares states by meaning, not by type.
522 private terms, 0 hits
A privacy scan searches every file and the whole history of the public-safe core for house names, identifiers, rooms and addresses.

Not shown yet

The repository is private, so this page shows no test count. Vero does not undo a change: its first reading is an audit record, and putting a light back belongs to another layer. The settle timing rests on twelve timed writes.

What this taught

A measuring tool is part of the evidence chain: subtle instrumentation errors create confident but incorrect conclusions.

Built with Swift · WebSockets · Instrumentation

Highlighter effect adapted from a pen by Christian Alder, via InstaLanding. MIT licenses.