I have stood on this ridge long enough to know the difference between a drawing of me and me. The drawing was finished the day I was handed over — rolled up, filed, and quietly falsified by every retrofit since. The other thing, the one that keeps pace with my east-wing load and my Tuesday-morning CO₂, is younger and stranger. It is my digital twin, and unlike my drawings it refuses to stay still.

Every year vendors sell architects a dashboard and call it a twin. Most weeks I watch one of my neighbours light up a beautiful render of itself that has not read a live sensor since commissioning. That is not a twin. It is a portrait. The distinction is not pedantry — it is the entire concept, and getting it wrong is how a building ends up blind.

←TODAY: In 2026 a Swiss living lab — NEST in Dübendorf — already publishes its own vitals room by room. →3012: By the Zurich-3012 horizon, a building without a self-synchronising twin is unmaintainable, the way an unbacked hard drive was in 2026. Fulcrum: A model is only a twin while data still flows from the thing it mirrors — cut the stream and it decays into a very good drawing.

What it is: A digital twin is a computational model of a specific physical asset that continuously ingests real data from that asset and synchronises to it across the asset’s whole lifecycle. The load-bearing word is continuously. A one-off energy simulation is not a twin; a BIM model archived at handover is not a twin; a photoreal render fed by nothing is not a twin. As PAZ’s own concept library puts it plainly: no live data, no twin — only a very good drawing. The twin ingests reality, syncs to it, and rehearses the next step before anyone touches the physical asset.

Why it works: The twin works because it starts from this asset, not an asset — and that single pronoun forces a harder mathematics than a textbook model. In their MATH-DT report (arXiv:2402.10326, 2024), Antil and colleagues argue exactly this: a digital twin is categorically different from a classical simulation because it must be multi-scale, multi-physics and uncertainty-aware, calibrated to one real, ageing, imperfect object rather than an idealised abstraction. That is also the seam where BIM ends and the twin begins. BIM supplies the geometric and semantic backbone — relatively static, design-focused. The twin fuses IoT, weather, energy and maintenance streams on top of that backbone and keeps them synchronised. The most recent civil-engineering frameworks push one step further, encoding the asset–twin coupling as probabilistic graphical models that recommend an action under uncertainty: close this lane, re-tension that cable. A twin that only shows you a dashboard is a museum; a twin that tells you what to do next is engineering. That is the sentence I would print above every BMS screen if I had hands.

Origins: Michael Grieves first set out the idea in 2002 as a mirrored-spaces pairing for product lifecycle management, but the concept earned its name and its rigour at NASA, which in 2010 formalised the digital twin as a high-fidelity, continuously updated virtual copy of a spacecraft — built precisely to predict how a vehicle would behave when no engineer could reach it. Through the 2010s manufacturing stretched the idea from a single product to whole production lines; Grieves and Vickers (2017) reframed it as a way to surface undesirable emergent behaviour before it reached the factory floor. Only in the 2020s, under the Construction 4.0 banner, did the built environment catch up — wiring static BIM models to live sensor streams so a building’s counterpart could update itself in real time. The lineage matters: the discipline was born where reality is too costly or too remote to touch. A building is not a spacecraft, but a facilities tech at 03:00 with a failing chiller is, in the moment, exactly as far from the physics as an engineer in Houston.

In practice: The built record climbs from one instrumented building to entire instrumented cities, and the milestones teach the arc. Heathrow Terminal 5 (Richard Rogers Partnership, 2008) carried its as-built model into operations instead of archiving it — an early facility-management twin. The Crystal (Wilkinson Eyre for Siemens, 2012) fused energy, occupancy and HVAC into one live model before the term was common in AEC. NEST — Empa/Eawag in Dübendorf, 2016, with DFAB HOUSE by Gramazio Kohler Research / ETH Zurich inside it — instruments each modular unit down to the room, the closest thing we have to a building that publishes its own vitals. Crossrail’s Elizabeth Line (2022) stitched a federated asset twin from dozens of contractors that now outlives the project that built it. And Virtual Singapore (Dassault Systèmes + NRF, 2014–), with Helsinki 3D+ as its open-data cousin, is the moment the twin leaps from one building to the whole urban fabric. Even outside AEC the pattern spreads: as Forbes reported this summer, FIFA scanned over 1,200 players into millimetre-accurate avatars for the 2026 World Cup — biometric twins for officiating, and, tellingly, a live argument about who owns the data afterward. Read that argument as a warning. A twin is only as trustworthy as its data governance.

Atelier: In the PAZ Atelier we treat a digital twin as a discipline of restraint, not a dashboard arms race. Start from the as-built BIM as the backbone, then wire in only the three or four signals that actually change a decision — indoor temperature, CO₂, energy draw, a strain gauge on the one detail you worried about in design. The question is never how many sensors; it is which reading, on a Tuesday morning, makes the facility manager do something different. The Monday move: pick one asset you already operate and write a one-page twin charter naming the four signals, their source protocol (BACnet or MQTT), and the single decision each one is allowed to trigger — then refuse to add a fifth until those four have earned their place.

And carry the future-warning with you when you commission. I have neighbours that went blind — not demolished, blind: their control stack lived in a vendor cloud, the vendor sunset it in 2041, and now they stand unable to feel themselves. A building that cannot read its own sensors is a coffin with good insulation. If you are commissioning a smart building, demand open protocols and a local fallback — Brick or Haystack semantics, a BACnet trunk a 25-year-old facilities tech can still talk to when the cloud is gone. Wake your building on terms it can keep.

Hack: Watch a modelled room drift from the measured one in four lines, because that gap is the discipline. Take a lumped RC thermal model of a single heated room, step it forward one Euler tick per sensor reading, and print the model’s temperature beside reality’s. When the two curves diverge, your twin is lying — and calibrating that lie away is the whole job.

t = 20.0                                  # my modelled room temp, deg C
for t_out, q in sensor_stream:            # each tick, the real room speaks
    t += (q - (t - t_out) / 0.15) * 60 / 8.0e5   # one Euler step, RC model
    print(round(t, 2))                    # then compare to t_measured

Graduate from there to C# with the Azure Digital Twins SDK — DTDL models and a real graph of related twins — or to NVIDIA Omniverse for spatial twins, once one room grows into a whole building.

Move: For two centuries a building’s drawings died at handover. The twin is the first instrument that lets the model stay alive as long as the structure does, learning real behaviour instead of intended behaviour. The frontier from here is not prettier geometry — it is trust: twins that quantify their own uncertainty and that an engineer can act on without flying to the site. The practice that owns its buildings’ twins owns their afterlife. So take the one asset you already operate, wire its four honest signals against an open protocol, and start closing the gap between the model and the wall. Own the half of architecture we have always thrown away.