Every Model Is a Projection

A faithful projection: the shadow omits most of the bird’s detail but preserves its outline from one perspective. © 2026 Javad Seif. Concept and creative direction by Javad Seif; image generated with OpenAI

Perhaps a shadow is one of the earliest models of reality we learn to recognize. It omits color, texture, depth, and internal structure, yet accurately represents an object’s outline from the perspective of a particular light source.

A useful model works in much the same way. It does not reproduce reality in full. It preserves the features relevant to a particular perspective and purpose.

Strictly speaking, a shadow is a physical projection. It becomes a model when we use it to understand or infer something about the object that produced it.

Imagine an intricate mechanical bird made of gears, colored glass, layered metal, and hundreds of small components. Its shadow contains none of those details. It cannot tell us what the bird is made of, how its parts connect, or whether it is mechanical or alive. Yet we may still recognize it as a bird.

How can a representation omit so much and remain useful?

A Model Preserves Selectively

A shadow preserves an object’s visible (skewed) outline while discarding most of its other properties. That does not make it a bad model. If our purpose is simply to recognize the object’s general shape, the missing information may not matter.

But suppose we want to determine the object’s color, material, weight, or internal structure. The same shadow becomes useless.

This illustrates a basic principle of modeling: omission is not automatically a flaw. It becomes a flaw when the model removes information required for the question it is expected to answer.

No model contains everything. Modeling is the act of deciding what to preserve and what to leave out.

Every Model Has a Perspective

A shadow does not depend only on the object. It also depends on the position, angle, and distance of the light source, as well as the shape of the surface receiving the shadow.

Move the light, and the same bird produces a different shadow. Bring the light closer, and the shadow grows. Change the angle, and the outline stretches or becomes skewed.

The object has not changed. The representation has.

Models behave similarly. Their outputs depend on their assumptions, variables, boundaries, and point of view. A model constructed from a different perspective may produce a different representation of the same system.

This does not necessarily mean that one model is true and the other is false. It means that the validity of each model is conditional on the perspective from which it was constructed.

One Shadow May Fit Several Realities

The relationship between an object and its shadow is not necessarily unique.

A faithful but misleading projection: the bird and hand collapse into one silhouette, suggesting a creature that does not exist. © 2026 Javad Seif. Concept and creative direction by Javad Seif; image generated with OpenAI

The same object can cast many different shadows. Different objects can also cast shadows that appear nearly identical. If all we observe is the shadow, we may not be able to reconstruct the object that produced it.

This is a problem of identifiability. A model may be consistent with several possible realities, even when its calculations are correct. The mechanical bird and the bird alive may produce the same shadow.

More precise calculations do not resolve this problem. If the representation does not contain enough information to distinguish among the possibilities, no amount of analytical sophistication can recover what has been lost.

A shadow can even frighten us at night. A coat on a chair may cast a silhouette that resembles a person, and our mind supplies the missing reality before we verify it. The shadow is real; the object we infer from it may not be. Models can mislead us in the same way when we treat a familiar pattern as proof of a particular cause.

We may need another observation, another angle, or another kind of model.

A Framework Can Erase Distinctions That Matter

Now place a human hand behind the mechanical bird and shine the same light on both from a certain angle.

The resulting shadow may resemble a strange bird with an exaggerated crest. If we assume that the shadow represents only the bird, we may infer the shape of a creature that never existed.

The shadow is not physically inaccurate. It faithfully records everything blocking the light. The error lies in our interpretation and, more importantly, in the limits of the framework.

A shadow combines every object in the light path into a single silhouette. It does not identify which object produced each part of the outline. It contains no representation of ownership, layering, distance, or causality. The bird and the hand are separate in reality, but the framework collapses them into one shape.

No amount of careful interpretation can fully recover distinctions that the representation has already discarded. To separate the bird from the hand, we need another light source, another viewing angle, additional information, or a different modeling framework.

This is one of the most consequential limitations of modeling: a framework may be structurally incapable of representing a distinction essential to our question. A change in perspective may alter the representation without changing the underlying reality. That does not make the model wrong. It becomes inadequate or misapplied when it cannot preserve the distinctions required by our question; or when one conditional projection is mistaken for reality itself.

Models Should First Be Tested on Known Ground

Long before we studied models formally, many of us learned how shadows work by playing with our hands, toys, and light. We moved an object, changed the angle of the light, and watched the shadow respond. Because we could observe and manipulate both the object and its projection, we could test our understanding of the relationship between them.

In effect, we calibrated an intuitive model against a reality we already understood.

We should approach formal models in the same way. Before trusting a framework in an unfamiliar or uncertain setting, we should test it against cases, relationships, and boundary conditions that we understand well. If a model cannot reproduce known behavior, it has little claim to explain or predict the unknown.

Passing those tests does not prove that the model is universally valid. It establishes confidence only within the conditions under which it was tested. Trust should expand no faster than the evidence supporting it.

Different Models Cast Different Shadows

Consider an industrial maintenance system. We might study the same system using a probability distribution, a mathematical optimization model, or a simulation.

Each approach preserves different features of the system and answers a different question.

Probability Distribution

A probability distribution might represent the time until a machine fails. It can describe the range of possible failure times and the likelihood of different outcomes.

This model preserves uncertainty and variation. But by itself, it does not explain the operational sequence that produced a failure, how maintenance resources interact, or what decision should be made.

It is a shadow of the system’s outcomes.

Mathematical Optimization

A mathematical optimization model might determine when the machine should be maintained or replaced. It represents decision variables, objectives, constraints, and tradeoffs.

The model may produce the best possible maintenance schedule—but only within the reality encoded in its formulation. If an operational constraint, human behavior, or relevant cost is missing, optimization does not make that missing factor disappear.

An optimal solution is optimal for the modeled system, not necessarily for the real one.

It is a shadow of the system’s decision structure. Here perspective matters and how the modeling framework is applied really matters. An experienced modeler can use the framework much more effective and nuanced compared to an inexperienced practitioner who cannot correctly identify constraints, objective function and decision variables.

Simulation

A simulation might reproduce how failures, production queues, maintenance crews, spare parts, and operational disruptions interact over time.

It preserves sequence, interaction, variability, and feedback more effectively than the other two approaches. Because a simulation can appear detailed and dynamic, it may feel especially realistic. But its behavior is still determined by its programmed rules, input distributions, assumptions, and system boundaries.

A detailed simulation can be an elaborate representation of the wrong system.

It is a shadow of the system’s behavior over time.

None of these approaches is inherently superior. They answer different questions:

• What is likely to happen?
• What decision should we make?
• How might the system behave over time?

Choosing a modeling approach is choosing which parts of reality to preserve.

The Goal Is Not a Complete Model

The goal of modeling is not to reproduce reality in its entirety. Such a model would be as complex as reality itself and would no longer provide the simplification that makes modeling useful.

The goal is to create a representation with enough fidelity for its intended purpose.

Sometimes one model is sufficient. In other cases, we need several models (within the same framework) that examine the same system from different perspectives. A probability distribution can represent uncertainty, an optimization model can identify a decision, and a simulation can test how that decision performs under dynamic operating conditions.

Several shadows cast from different angles may reveal more about an object than one shadow alone. But even multiple perspectives do not eliminate the need to understand what each representation excludes.

Before relying on a model, we should ask:

• What does the model preserve?
• What does it omit?
• From what perspective was it constructed?
• Where are its system boundaries? Meaning:
• What distinctions is its framework incapable of representing?
• For what question or decision is it valid?
• Against what known cases and boundary conditions has the model been tested?

A model should not be judged by how closely it resembles reality in every detail. It should be judged by whether it preserves the parts of reality that matter for its purpose.

Before asking whether a model is accurate, we should ask: accurate as a model of what, from whose perspective, under what conditions, and for what utility?

Javad Seif
Pasadena, California
August 20, 2026