Engineering & Technologypreprint2026-09-27

The Swarm House Reference Architecture: HouseBench — a benchmark for adaptive physical intelligence — eleven layers, forty-six classes, one reversible interface, seven architectures, fifteen adversarial tasks and twelve procedural houses

Open access0 citations

Abstract

Intelligent-building research has hardware, and it has decision layers, and almost no way of telling whether a given architecture keeps a dwelling habitable once the dwelling changes. This paper supplies both halves. The first is a reference architecture, stated completely enough to implement: eleven layers, an ontology of 46 machine-readable classes each carrying seven fields, an entity model whose tenth field is provenance, one reversible interface with 5 channels and an 8-question handshake, perception measured eight separate ways, an 11-step runtime loop with the safety filter inside it, and eight invariants below everything that reasons. The second is HouseBench: 12 procedurally generated dwellings, seven architectures, 15 adversarial tasks, a removal ladder of 6 rungs, 16 quantities measured separately, and 882 logged runs. The headline result is uncomfortable for everybody, this author included. With nothing broken every architecture above a commissioned table performs alike. Separation arrives under injury, and it is separation by dependence: the centre removed takes three architectures from 38 per cent of the day to 0 per cent; carriage cut takes the coordinated system to 0 per cent; and a confident wrong reading leaves six of seven at 8 per cent against 38 per cent for the one carrying provenance and a command log. On eight plans nobody tuned against, a commissioned table holds 0 per cent. Safe occupancy holds in 315 of 315 task runs and at every rung, because the filter is a step of the loop and the release is a spring. Swarm Room 0 closes the paper at 11 995 euro, and so does a negative finding: an architecture shown working on a healthy house has been shown almost nothing. What this paper settles – The architecture is complete enough to build: 46 classes, 17 relations, 83 capabilities, 77 constraints, 156 failure modes, one joint family, eight invariants. – On a healthy house, seven architectures are indistinguishable above a commissioned table. – Dependence is measurable, and it is the whole of the difference. Centre removed: 0 per cent against 38 per cent. Carriage cut: 0 per cent against 38 per cent. – Provenance earns its field. Under a reading no command asked for: 38 per cent against 8 per cent. – Generalization is the cheapest test and the most decisive. A commissioned table holds 0 per cent on eight unseen plans. – Intelligence and safe occupancy come apart: 315 of 315 task runs held it, and so did every rung of the ladder. – Seven ablations locate the advantage in one part; five of the other six earn nothing measurable here. – A first build is specified and affordable: 20.25 square metres at 11 995 euro, the reasoning inside that 349 euro. Author ORCID: 0009-0008-6255-7724. License: Creative Commons Attribution 4.0 International.

// Source

View paper (DOI)Open access versionOpenAlexZenodo (CERN European Organization for Nuclear Research)Published 2026-09-27

Authors: Gonçalo Melo de Magalhães