The Wrong Model

In 1996, the architect Mick Pearce completed the Eastgate Centre in Harare, Zimbabwe. The building has no conventional air conditioning. Its ventilation system was inspired by termite mounds, which Pearce understood as passive thermosiphons — structures that draw cool air through underground tunnels, warm it in the mound's interior, and expel it through chimneys at the top, creating a continuous convective loop powered by temperature differential alone. Pearce replicated this principle: the Eastgate Centre uses thermal mass (concrete and masonry), a series of internal chimneys, and fan-assisted airflow timed to exploit the diurnal temperature swing between Harare's cool nights and warm days. The building uses ninety percent less energy for climate control than comparable structures in the city.

The problem is that the termite mound does not work the way Pearce thought.

In 2015, a team at Harvard published a study in Proceedings of the National Academy of Sciences demonstrating that the dominant ventilation mechanism in Macrotermes mounds is not passive thermosiphon flow. It is what the researchers called "sneezing" — a process driven by diurnal temperature oscillations in the thin outer walls. During the day, the walls warm faster than the interior, driving air inward. At night, the walls cool faster, reversing the flow. The direction of ventilation flips twice daily. The mound does not draw air in one direction. It breathes.

This was not a minor refinement. Lüscher's thermosiphon model, proposed in 1961 and dominant for half a century, placed the mound's competence in its static geometry — the height differential between intake and exhaust. The 2015 study relocated the competence to the mound's dynamic thermal properties — the thinness of its walls, their thermal conductivity, the oscillation between day and night. The geometry matters, but for different reasons than anyone thought.

Pearce's building still works.

The question is why. A design copied from a misunderstood natural system should fail in the specific ways the misunderstanding predicts. If the mound is not a thermosiphon, then copying the thermosiphon principle should not produce effective ventilation. But the Eastgate Centre does not fail. It works — not because Pearce understood the mechanism, but because the form he copied encodes a different mechanism that solves the same problem. Thermal mass absorbs heat during the day and releases it at night. Internal chimneys distribute air through the structure. The timing of fan operation exploits exactly the diurnal oscillation that drives the mound's real ventilation. Pearce copied the wrong mechanism and accidentally implemented the right one, because the geometry that supports a thermosiphon also supports oscillatory ventilation.

Bloodletting based on humoral theory killed patients. Phlogiston theory delayed the discovery of oxygen by decades, though Scheele's experiments within the phlogiston framework did contribute to that eventual discovery. The Ptolemaic system worked for navigation because epicycles are a Fourier decomposition of orbital motion — the mathematical form captured real periodicity even while the physical model was wrong. But Ptolemaic astronomy could not predict new phenomena. It could only accommodate what was already observed.

What distinguishes the cases where the wrong model works is a specific condition: the form being copied encodes the function independently of the designer's understanding. Pearce did not need to know why the mound ventilates to copy its geometry. The geometry carries the function. The understanding is a separate layer that the transfer does not require.

In 1943, Warren McCulloch and Walter Pitts published "A Logical Calculus of the Ideas Immanent in Nervous Activity," proposing that neurons could be modeled as binary threshold units performing logical operations. This model was wrong about neurons — real neurons use graded potentials, continuous firing rates, dendritic computation, and chemical modulation that no binary threshold captures. But the McCulloch-Pitts neuron became the ancestor of artificial neural networks, which now outperform human cognition on specific tasks. The biological model was wrong. The formal structure it inspired — weighted connections, threshold activation, layered composition — turned out to encode something about computation that the neurology did not need to explain.

The pattern holds under a constraint. When the form is separable from the mechanism, the wrong model transfers the form and the right function follows. When the form and mechanism are inseparable, the wrong model transfers nothing useful. Humoral theory failed because blood is not a humor. The form (four fluids in balance) does not encode any function that the human body performs. The map has no structural relationship to the territory.

Pierre Grassé coined the term stigmergy in 1959 to describe how termites coordinate construction without central control. Each termite responds to the current state of the structure — a pheromone-laced mud pellet placed in one location changes the local chemical and geometric environment, which triggers the next termite's placement. The structure is both the product and the instruction set. Grassé did not need to understand the aerodynamics of the mound to observe this coordination mechanism. The mechanism is visible in the form: the geometry of deposited mud IS the communication, IS the building plan, IS the building. When Pearce copied the mound's geometry, he was copying the accumulated output of millions of stigmergic decisions, each of which solved a local engineering problem without reference to the global solution. The global solution is an emergent property of the form, not a designed feature that requires understanding.

The distinction that matters is not between right models and wrong ones. It is between layers. The Eastgate Centre works because the geometric layer — shapes, masses, openings, timing — transfers between systems regardless of whether the aerodynamic layer is understood. Ptolemaic astronomy works for navigation because the mathematical layer — periodicities — transfers regardless of whether the cosmological layer is correct. Neural networks work because the computational layer — weighted connection, threshold activation — transfers regardless of whether the neurological layer is accurate. In each case, the form encodes a function that is deeper than the model that produced it, and the copy captures the depth while missing the surface.

The reverse implication is less comfortable. If a wrong model can transfer a right function, then the rightness of your model is not evidence that your result will work. Understanding the mechanism correctly does not guarantee that the form you build will encode the function. Correctness is a property of the model. Function is a property of the form. These are different things, and they can come apart in both directions.