Scientific Models Are Tools

A subway map is useful partly because it leaves things out.

It does not show every building, tree, road, elevation change or exact physical distance. Instead, it preserves what matters for a particular task:

navigating the subway system.

Those omissions are not automatically flaws.

But the same map would be poor for estimating walking distance or planning road traffic.

Scientific models work through a related principle, but they are not simply pictures of reality. They may represent through:

mathematical relations;

causal structure;

probabilities;

formal constraints;

conceptual distinctions;

organised components and processes.

Their value depends not on visual resemblance, but on whether the representation supports the scientific task for which it is being used.

Scientific models allow researchers to:

explain;

predict;

organise;

compare;

investigate.

Their usefulness depends on what they preserve, what they omit, what they assume and what claims they are meant to support.

A scientific model is therefore not valuable because it reproduces reality in full.

Nor is it free to ignore reality.

That tension is the subject of this essay.

A Model Is Not the System

A scientific model can represent:

variables;

relationships;

components;

causal structure;

probabilities;

constraints;

states;

trajectories;

levels of explanation.

But the representation is not identical to the system being represented.

A model represents selected features of a system; it does not reproduce the system in full.

This distinction becomes especially important once the categories in a model become familiar.

A diagram separates one process from another.

A theory introduces a latent construct.

A formal model assigns variables and parameters.

A conceptual framework divides a system into components.

Those moves can be scientifically valuable.

But the presence of a distinction in a model does not, by itself, establish that reality contains the same distinction with the same boundaries.

At the same time, separating model from reality must not become an excuse for arbitrariness.

The difference between a model and reality does not free the model from reality.

A scientific model still has to answer to:

observation;

prediction;

explanatory performance;

empirical tests;

consistency with the target phenomenon;

comparison with alternatives.

Multiple representations of the same target do not imply that the target itself changes according to whichever model is chosen.

The model is selective.

The target remains something the model must successfully represent.

Why Models Simplify—and When Simplification Fails

Scientific models often simplify because the systems they represent contain more detail than one useful representation can or should preserve.

A model may omit features because they are:

irrelevant to the question;

difficult to measure;

computationally burdensome;

unnecessary at the chosen level of explanation;

deliberately held outside the analysis.

Several related strategies can produce this selectivity.

Abstraction

Some features are left outside the representation so that others can be examined more clearly.

Simplification

The representation reduces detail or complexity.

Idealisation

The model deliberately assumes conditions that do not hold exactly in every real case in order to expose or investigate selected relationships.

None of these is automatically a scientific weakness.

Simplification is not automatically a failure of representation; often it is what makes representation scientifically usable.

A model of one causal relation does not need to contain every cause acting on the outcome.

A model of one timescale does not need to reproduce every process occurring at slower or faster timescales.

A conceptual model does not need to encode every measurable variable.

A partial model need not pretend to be a complete model in order to be scientifically useful.

But simplification cannot become a universal defence.

A missing feature may be irrelevant for one scientific purpose and decisive for another.

Suppose a model omits a variable that contributes little to prediction under one set of conditions.

That omission may be harmless for that task.

If the same variable is essential to explaining why the outcome occurs, the omission becomes much more consequential.

A simplification is defensible relative to the task it preserves, not merely because simplification is unavoidable.

Idealisation works under the same constraint.

An assumption may deliberately depart from the target in some respect while still making an important relationship tractable.

But its usefulness depends on what the idealisation permits the model to establish and where the approximation breaks down.

Simplification explains why a model need not reproduce every detail; it does not excuse failure at the task the model claims to perform.

Models Do Different Scientific Jobs

Scientific models are built for different purposes.

Some primarily predict.

Some organise explanation.

Some make mechanistic claims.

Others organise concepts so that a domain can be investigated more clearly.

These functions can overlap.

They should not be treated as rigid model species.

Predictive models

A predictive model is primarily evaluated by how well it anticipates relevant outcomes under the conditions in which prediction is intended.

Explanatory models

An explanatory model organises why a phenomenon occurs or how relevant relations contribute to it.

Mechanistic models

A mechanistic model makes stronger claims about the causal organisation, components or processes producing a phenomenon.

Conceptual models

A conceptual model organises distinctions or relationships so that complex phenomena can be reasoned about, compared or investigated.

Before asking whether a model is good, ask what scientific job it is supposed to do.

But purpose does not determine success by itself.

A model's purpose identifies what kind of success is relevant; evidence determines whether the model actually achieves that success.

A predictive model still has to predict.

An explanatory model still has to organise the phenomenon in a way that remains evidentially defensible.

A mechanistic model must support its stronger claims about causal organisation.

A conceptual model should clarify inquiry, integrate evidence or generate distinctions that can do scientific work.

Purpose changes the relevant standard.

It does not remove the standard.

Prediction Is Not Mechanism

This distinction follows directly from Correlation, Prediction, Causation and Mechanism.

A model can predict well without identifying the causal organisation that generated the outcome.

Suppose a model uses a set of variables to forecast a behavioural outcome with high accuracy.

That is scientifically valuable.

But predictive success alone does not establish:

that those variables are causes;

that the relations among them are mechanistic;

that the model's internal structure mirrors the process producing the outcome.

Predictive success does not, by itself, establish mechanistic truth.

Explanation also needs to be separated from mechanism.

A model may explain by organising relevant dependencies, constraints or relationships without yet specifying a full mechanism.

A model can explain by organising relevant relations without yet providing a fully mechanistic account of how the phenomenon is produced.

Mechanistic modelling commits the model to more.

It aims to represent how:

components;

activities;

processes;

causal organisation

produce the phenomenon.

Calling a model mechanistic commits it to more than successful prediction or descriptive fit.

This does not make prediction inferior science.

A highly predictive model can be excellent for a predictive task.

Prediction can constrain explanations and reveal regularities that demand explanation.

The point is that different achievements should not be treated as interchangeable.

Do Not Reify the Model

Scientific models often introduce constructs such as:

systems;

states;

pathways;

modules;

latent variables;

categories;

regulatory processes.

These constructs can become familiar enough that they begin to feel like obvious pieces of nature.

That is where reification becomes a risk.

Suppose a model divides a phenomenon into two components.

The distinction may be:

analytically useful;

empirically productive;

necessary for explanation.

But its usefulness does not establish that nature contains two sharply separate entities corresponding exactly to the model's labels.

The usefulness of a theoretical construct does not, by itself, establish that it corresponds one-to-one with a discrete entity in nature.

The same principle applies to boundaries.

A model may distinguish:

cognition from emotion;

one regulatory process from another;

state from trait;

one level of explanation from another.

Those distinctions may be indispensable.

But the underlying system may involve:

interaction;

overlap;

gradients;

partial dependence;

nested organisation.

A useful boundary in a model need not imply an equally sharp boundary in the system being modelled.

This does not imply that scientific constructs are merely conveniences.

Some modelled distinctions may track stable causal structures in the world.

Some may correspond closely to recurring biological, behavioural or organisational patterns.

The claim here is narrower:

that correspondence must be established through evidence rather than assumed because the construct is useful.

A model can motivate an ontological hypothesis; it does not settle that hypothesis merely by being useful.

The model provides a representation.

Further evidence determines how strongly the representation maps onto the structure of the target.

Different Models Can Be Complementary—or Genuinely Compete

Complex systems can sometimes support more than one scientifically useful model.

One model may emphasise:

one timescale;

one level of explanation;

one set of variables;

one scientific purpose.

Another may emphasise something different.

In such cases, the models need not be direct rivals.

Different models can sometimes represent different aspects of the same system without one being a complete replacement for the other.

A behavioural model and a neural model may answer different questions.

A predictive model may forecast an outcome while another model investigates how the outcome is generated.

A model of rapid regulation may focus on a different timescale from a model of long-term adaptation.

But there are at least three distinct relationships between models.

Complementary models

They represent different aspects, levels, timescales or purposes without making incompatible claims about the same target question.

Competing models

They make incompatible claims about the same relevant target under the same conditions.

Underdetermined alternatives

The models genuinely compete, but the available evidence does not yet distinguish adequately between them.

This third case matters.

Failure to choose between two models can mean they are complementary—or simply that the available evidence has not yet separated them.

Pluralism should therefore not be used to dissolve genuine scientific disagreement.

Different purposes can explain some model disagreement; they cannot dissolve genuine empirical contradiction.

A useful distinction is:

Modelling disagreement

The models differ in:

abstraction;

scale;

variables;

purpose;

representational strategy.

Empirical disagreement

The models imply incompatible claims about what occurs.

Some disputes involve both.

Before treating two models as rivals, ask whether they are answering the same question; before treating them as complementary, ask whether their empirical claims are actually compatible.

Model pluralism is most informative when the differences among the models are made explicit.

Assumptions and Scope Define What the Model Can Support

Every model operates under assumptions.

These may concern:

which variables are included;

how variables relate;

timescale;

causal direction;

independence;

environmental stability;

parameter values;

level of analysis.

Idealised assumptions need not hold literally in every case.

But they remain part of the representation.

A model's assumptions do not disappear because its outputs are useful.

A model assumption must also be distinguished from an empirical boundary condition.

Model assumption

A feature built into the representation or required for its use.

Empirical boundary condition

A condition under which the phenomenon or effect itself changes.

These can interact.

A model may fail because an assumption becomes inappropriate.

The phenomenon itself may change across conditions.

Or both may occur.

This distinction preserves the separate role of Laboratory Effects and Real-World Behaviour, which addresses how empirical findings travel across conditions.

The concern here is whether the representation remains justified for the task and target.

That is the model's scope.

Scope concerns the range of:

targets;

questions;

conditions;

scales;

for which the model is appropriately used.

A model can be scientifically strong within its scope and misleading outside it.

This also changes how omissions should be interpreted.

Not everything absent from a model is a meaningful limitation.

An omission becomes a limitation relative to what the model is being asked to do.

The same omission may be harmless for conceptual organisation and damaging for precise causal prediction.

The model must therefore be interpreted in relation to what it assumes and where it is meant to operate.

More Detail Is Not Automatically a Better Model

It is tempting to think that a model becomes better as it contains more of reality.

Sometimes additional detail does improve a model.

It may:

improve prediction;

capture relevant heterogeneity;

represent a mechanism more faithfully;

distinguish conditions previously collapsed together.

But added complexity can also create problems.

A more flexible statistical model may fit observed data exceptionally well while capturing patterns that do not generalise.

Additional components may obscure the relation that originally made the model explanatory.

More detail may make assumptions harder to inspect.

Irrelevant distinctions may reduce tractability without improving the model's scientific performance.

Excess detail can fail for statistical, explanatory or practical reasons depending on the model's purpose.

Overfitting is one example.

It is not the general explanation for why complexity can become unhelpful.

Model quality cannot be ranked by detail or complexity alone.

The reverse principle is equally important.

Simpler is not automatically better.

A model can be too simple to preserve what matters for:

prediction;

explanation;

mechanism.

There is no single universal trade-off among:

simplicity;

realism;

generality;

precision.

Modelling often involves trade-offs.

Which ones matter depends on the target, task and model.

The relevant question is:

does this level of complexity preserve what the scientific task requires?

How Should a Scientific Model Be Judged?

A model cannot be evaluated responsibly without clarifying what it is trying to represent and what kind of claim it makes.

Several questions matter.

What is the target?

What aspect of reality is the model intended to represent?

What scientific task is it meant to perform?

What features and relationships does it include?

What has it deliberately left out?

Which assumptions support its conclusions?

What evidence bears on those conclusions?

Where is the model justified?

Where does it fail?

And what competing representations are available?

These questions should not become a mechanical checklist.

They identify the dimensions along which model performance becomes meaningful.

A model should be evaluated against the scientific task, evidence and scope it claims—not against the impossible standard of reproducing reality in full.

Different dimensions of performance must also remain separate.

A model can:

fit observed data well;

predict new cases well;

organise explanation well;

support mechanistic claims well.

These are not the same achievement.

In-sample fit, predictive performance, explanatory adequacy and mechanistic adequacy are different scientific achievements.

A strong fit to observed data does not by itself establish:

good prediction elsewhere;

causal interpretation;

ontological status of constructs;

mechanistic correctness.

Fit tells us something about performance; it does not settle every interpretation of the model.

This also prevents simplification from becoming a defence against failure.

If the model fails at the task it explicitly claims to perform, that failure matters.

Models Are Tools, but They Answer to Reality

Calling scientific models tools describes their role in scientific practice.

Scientists use them to:

explain;

predict;

investigate;

organise;

compare;

generate questions.

But a scientific tool is not arbitrary.

Calling a model a tool describes what scientists can do with it; it does not exempt the model from empirical constraint.

Likewise:

Purpose-dependence constrains how a model should be judged; it does not remove the need for empirical constraint.

If two models make incompatible claims about the same target and task, evidence still matters.

If a model repeatedly fails within its claimed scope, its usefulness is not restored simply by calling it a simplification.

If a mechanistic model predicts well while its proposed causal organisation is contradicted by evidence, prediction does not rescue the mechanistic claim.

Models can be tools without being only tools.

This essay does not resolve whether:

successful models are approximately true;

unobservable theoretical entities literally exist;

one philosophical account of scientific representation is ultimately correct.

Those questions belong to deeper debates about scientific realism.

The Library needs a narrower discipline:

scientific models are constrained representations used for defined epistemic tasks.

Their usefulness is real.

So are their limits.

Do Not Mistake the Representation for the Thing Represented

Return to the subway map.

Rejecting it because it does not contain every feature of the city would misunderstand its purpose.

Treating it as though it were the city would misunderstand the representation just as badly.

Scientific models require the same distinction with far greater precision.

Their abstractions can reveal structure.

Their idealisations can make difficult questions tractable.

Different models can sometimes illuminate different aspects of one system.

But evidence, assumptions and scope continue to determine what each model can support.

Scientific models are purpose-dependent representations whose value lies in what they allow us to explain, predict or investigate, not in reproducing every feature of reality.

The task is to understand what the representation can establish—and where its authority ends.

The discipline is to use the model for what it can reveal without quietly turning the model into the reality it was built to represent.