A measurement-and-verification-firewalled living lab in an occupied single-family home

Abstract

Field demonstrations of predictive control in residential buildings report results but rarely how the multi-year evidence base was assembled, at what cost, or how the running controller was kept from contaminating the very measurements used to judge it. We describe an occupied single-family home being built into a research instrument - a multi-year programme whose instrumented phase began in 2026, standing on a 4.5-year measurement record (2022-) - under three simultaneous, mutually hostile contracts: a frozen measurement-and-verification (M&V) baseline, a live model-predictive controller (MPC) under continuous development, and a research programme of pre-registered experiments. The contribution is architectural: a three-plane firewall (Control / Analytics / Research) with explicit boundary rules, a documented failure-mode-and-effects analysis (FMEA) of the cross-plane corruption paths, and a set of operating patterns - a byte-hashed frozen baseline, a shadow-first / deploy-dark / kill-switch deployment lifecycle, a living incident register, an experiment-window masking invariant, and pre-registration of analyses before data collection - each of which closes a specific corruption path. We give logged evidence that the pattern bears load in practice: a code-review fix plan closed under live operation, a dead-on-arrival actuation bug caught in shadow and never in production, a specification assumption falsified by a natural experiment and the guard shipped the same day, a genuine cross-plane contamination - a control input locked in a self-confirming echo loop - exposed by an independent source and closed the same day, and a bit-identical model replication logged as a fact. The design is N = 1 by intent: it is a method paper about how to build one lived-in home into a credible instrument, not a performance claim about one house.