Models That Remake Reality

Every scientific model begins life modestly. It is a concession to limitation: too little data, too many variables, not enough time. A model simplifies so that action can occur at all. No physicist, economist, or epidemiologist mistakes a model for the thing modeled—at least not at first.

But something peculiar happens when a model leaves the lab and enters the world. It becomes embedded in decisions, incentives, and routines. It stops being merely a description and begins behaving like an instruction. At that moment, its status changes. The model ceases to be passive. It becomes a cause.

This transformation is the hidden machinery behind many of our most stable institutions and our most persistent errors.

The curious part is that the transition often goes unnoticed. Models slip from map to territory, not through persuasion, but through utility. They tell us what to measure, what to reward, and what to ignore. And once the machinery of coordination starts relying on them, the world begins adapting—sometimes gracefully, sometimes grotesquely—to their dictates.

The story is less about epistemology than engineering: how constraints propagate through systems, and how simplification becomes structure.

It is one of the minor ironies of modern life that our most “objective” metrics are often just outdated guesses trapped in a pension plan.

The part is straightforward: every model selects. It trims away detail—some irrelevant, some merely inconvenient. This is innocuous at the scale of personal understanding. A person can revise their beliefs when they encounter contradiction. A spreadsheet cannot.

When a model operates at institutional scale, its omissions harden into blind spots. What starts as simplification becomes filtration. People adapt their behavior to fit what is counted, funded, or sanctioned. Measurements turn into incentives, incentives into strategies, strategies into norms. The world’s degrees of freedom quietly shrink.

A prediction is made. People adjust. The prediction becomes true—not because the world naturally leaned that direction, but because we leaned on it.

This self-fulfilling dynamic is often mistaken for accuracy. In reality, it is the mechanical closure of a feedback loop.

Once a model provides stability—reducing uncertainty, standardizing decisions, enabling coordination—it can remain in place long after it stops representing anything interesting about the underlying system. Its survival depends not on fidelity, but on usefulness.

The whole is more consequential: as systems mature, they begin optimizing for the model rather than for the conditions that originally justified the model. Hospitals optimize for metrics; companies optimize for KPIs; universities optimize for rankings; financial firms optimize for risk forecasts that are based on data generated by the firms optimizing for the forecasts.

The model becomes the habitat. The original reality becomes the remainder.

And remainders accumulate. They take the form of corner cases, unmeasured costs, unmodeled incentives, and the people or processes that do not fit neatly into the simplified schema. These remainders exert pressure on the system. They create brittleness. They seed fragility.

A striking example comes from risk modeling in finance. Tools designed to quantify uncertainty gradually ended up shaping the very markets they were meant to analyze. Traders adapted, regulators adapted, products adapted. Over time, the models did not merely describe risk—they helped generate the conditions under which certain forms of risk became invisible, right up to the moment they returned dramatically in 2008.

The usual conclusion is that the models were wrong. But the more unsettling conclusion is that they were operative. They imposed a structure on behavior that was stable until it wasn’t. Their failure was not epistemic; it was ecological.

Similar dynamics appear in domains far from Wall Street. Educational rankings distort curricula. Policing algorithms reshape patterns of crime reporting. Productivity metrics redefine what “good work” means. Benchmarks become targets; targets become ceilings.

The mechanism is always the same: embed a model deeply enough in a decision process, and the system begins organizing itself around that model’s categories and incentives. Over time, the model becomes less a tool and more a constraint.

The implication is that fighting over whether a model is accurate can be profoundly beside the point. Accuracy matters only briefly—during adoption. After that, what matters is durability: whether the world can withstand being sculpted by that simplification.

A system optimized for a model is not optimized for the world. It is optimized for reducing the friction between the world and the model’s expectations.

This raises a practical question: if models inevitably reshape their domains, what distinguishes the models that support resilient systems from those that drive them toward collapse?

One answer is that robust models are modest. They treat their simplifications as negotiable. They allow error signals to penetrate. They preserve slack—degrees of freedom for the world to behave in ways the model did not anticipate.

Fragile models, by contrast, treat deviations as malfunctions. They force alignment. They eliminate slack. They produce a veneer of order that is actually a concentration of unreconciled pressures.

The difference between the two is not conceptual but architectural. It lies in how systems are built to incorporate—or suppress—feedback.

The temptation, when faced with these dynamics, is to reject abstraction entirely. But this would be an overreaction. Human life without models would be paralyzed. Coordination would vanish. Complexity would suffocate action.

The more sober conclusion is that every model carries a cost, and the cost rises with scale. The more people depend on a model, the more the world will be bent to its image. The danger arises not from modeling per se, but from forgetting that the world is the primary system and the model is the derivative.

Models are indispensable. But so is the ability to let them break cleanly rather than letting them break the world.

The challenge ahead is not to build perfect models, but to build systems that remain permeable to reality—systems that can absorb surprises, reincorporate remainders, and acknowledge the unmodeled without treating it as error.

The task is to design models that do not eat the world, but feed it: models that clarify without constraining, coordinate without deforming, and guide without governing.

That is not mysticism. It is simply engineering with the humility to remember which thing is the scaffolding and which thing is the building.

Available on Amazon

All my English books are ready for you in my shop: shop.willemdewit.work