One-sentence answer: A model works when it deliberately simplifies part of reality for a declared purpose, makes its assumptions and variables explicit, produces outputs that can be checked against the world, and is corrected when its predictions, explanations or decisions fail.
A model is not reality. It is a representation built to help someone do a job: explain a mechanism, estimate an unseen state, compare alternatives, predict what may happen, test a design, rehearse a decision or communicate a complicated system in a usable form. The moment we forget the job, a useful simplification can become a source of false certainty.
A good model is not the one that contains everything. It is the one that leaves out the right things for the question being asked—and tells us what it left out.
Quick Read: the causal chain
QUESTION / PURPOSE → REAL-WORLD TARGET → BOUNDARY → VARIABLES → ASSUMPTIONS → REPRESENTATION → PARAMETERS / INPUTS → CALCULATION OR REASONING → OUTPUT → TEST AGAINST OBSERVATION → ERROR / UNCERTAINTY → REVISION → USE WITHIN THE VALIDATED RANGE
The discipline is cyclical. A model that cannot be challenged by new observations is not functioning as a correctable model; it is becoming a story protected from the world.
1. Start with the job, not the model
The same system may need several different models because different questions preserve different features. A street map may preserve roads and junctions while ignoring building interiors. A weather model represents atmospheric state and dynamics at a chosen resolution. A school progress model may preserve assessment evidence, time and learning objectives while omitting countless details of a student’s life. A financial model may focus on cash flows and assumptions rather than physical operations.
Before judging a model, ask: What is it for? NIST describes a research model as a representation, pattern or mathematical description that can help replicate or describe a system, process or result; its research-data framework also distinguishes verification—whether the computational implementation represents the mathematical model—from validation—how accurately the model represents the real world for its intended uses.
This “for its intended uses” clause is crucial. A model can be good for one job and unsafe for another.
2. Models create boundaries
Reality contains more detail than any useful model can carry. Modelling therefore begins by drawing a boundary: what is inside, what is outside, what changes, what remains fixed, what is measured directly, and what is represented through a proxy or assumption.
- System boundary: what the model treats as the system.
- State variables: quantities or categories used to describe the system at a moment.
- Inputs: conditions supplied to the model.
- Parameters: values controlling relationships inside the model.
- Outputs: what the model returns.
- Assumptions: simplifications that make the representation workable.
- Operating range: where the model has evidence behind it.
Changing any of these can change what the model means. That is why a model transferred from one population, climate, machine, market, classroom or time period to another should not automatically inherit the same confidence.
3. A model selects relationships
Models become useful when they make relationships explicit enough to inspect. The relationship may be geometric, statistical, physical, logical, causal, behavioural or procedural. Some models are equations; some are diagrams; some are simulations; some are taxonomies; some are decision trees; some are physical prototypes; some are conceptual representations.
What matters is not whether the model looks sophisticated. What matters is whether its structure corresponds well enough to the target for the declared job.
4. Fitting a model is not the same as testing it
A model can describe existing data beautifully and still fail on new cases. NIST’s statistical guidance treats model selection, fitting and validation as distinct, iterative steps. That separation protects us from mistaking a close fit to known data for general reliability.
| Question | What it tests | Typical failure |
|---|---|---|
| Does it fit the data used to build it? | In-sample fit | Overfitting |
| Does it work on new observations? | Out-of-sample performance | Poor generalisation |
| Does the code implement the intended mathematics? | Verification | Implementation error |
| Does the model represent the real system well enough for the intended use? | Validation | Wrong structure or assumptions |
| Does the relationship remain when conditions change? | Robustness / transfer | Range failure or drift |
| Does it support the decision we want to make? | Decision relevance | Good model, wrong job |
5. Prediction is one test—not the whole definition of truth
Prediction is powerful because the model commits itself before the result is known. If the predicted state and observed state disagree, the difference becomes information. But a successful prediction does not automatically establish the model’s causal story, nor does one success validate every use of the model.
For example, a model may predict demand accurately without representing why people want the product. A statistical classifier may predict an outcome without identifying a causal mechanism. A weather model may forecast a storm using a vast dynamical system, while a simple empirical rule may perform adequately for a narrow local task. The right question is not “Which model is most impressive?” but “Which model is justified for this purpose, at this scale, with this evidence?”
6. Uncertainty belongs inside the model conversation
Inputs can be uncertain. Measurements can be noisy. Parameters can be estimated imperfectly. The model structure itself can be wrong. Future conditions may move outside the range represented in the data. These are different sources of uncertainty and should not be collapsed into one vague confidence score.
- Measurement uncertainty: uncertainty in observations or inputs.
- Parameter uncertainty: uncertainty about fitted values.
- Structural uncertainty: uncertainty about whether the relationships chosen by the model are adequate.
- Scenario uncertainty: uncertainty about future external conditions.
- Sampling uncertainty: uncertainty arising from the cases observed.
- Decision uncertainty: uncertainty about whether a model output supports a particular action.
A precise-looking output can therefore be misleading. The National Academies has repeatedly warned that model graphics and numerical precision can make results appear more certain than the assumptions and evidence justify. A world-class model report makes the uncertainty visible rather than hiding it behind decimals.
7. Model ≠ reality
| Model | A purposeful representation of selected features. |
|---|---|
| Reality | The system or phenomenon that continues to exist whether the model describes it well or badly. |
| Map | A particular spatial model designed for navigation, location or analysis. |
| Simulation | A model executed through time, events or repeated state transitions. |
| Prediction | An output about an unobserved or future state. |
| Explanation | An account of why or how a result occurs. |
| Evidence | Observations or records that can support, weaken or constrain claims. |
Keeping these distinctions clean prevents a common error: treating a model output as if the world itself had been observed.
8. Compare competing models, not only a favourite model against itself
A powerful hostile test is to ask whether another plausible model explains the same observations. Two models may fit existing data but diverge sharply under a new condition. Those divergent predictions tell us where to look next.
This is why modelling and observation form a loop rather than a one-way pipeline:
OBSERVE → MODEL → PREDICT → TEST → COMPARE → FIND ERROR → REVISE → OBSERVE AGAIN
9. Worked example: modelling a school commute
Suppose the question is: “What time should a student leave home to reach school reliably?” A weak model might use only straight-line distance. A better model might include walking time, bus frequency, transfer time, traffic, rainfall and a buffer. A still more detailed model could use real-time transport data.
But more detail is not automatically better. If the decision is simply whether leaving at 7:00 or 7:15 is safer, a modest model with reliable inputs may outperform an elaborate model dependent on stale or unavailable data.
The model should then meet the world: record actual departure and arrival times across several days. If its errors cluster on rainy mornings, the model has learned where its boundary is weak. The correct response is not to defend the model; it is to revise it.
10. The same mechanism appears across very different worlds
| Domain | Possible model | What must not be assumed |
|---|---|---|
| Physics | Mathematical representation of motion or fields | That every approximation works at every scale |
| Weather | Numerical atmosphere model | That one forecast is certainty |
| Engineering | Structural, thermal or system model | That simulation alone proves safe operation |
| Medicine | Risk or disease-progression model | That population estimates diagnose an individual |
| Education | Learning-progress or assessment model | That a score is the whole learner |
| AI | Statistical model producing classifications or generations | That fluent output equals verified truth |
| Economics | Model of flows, incentives or behaviour | That assumptions remain fixed after policy changes |
The transferable idea is the modelling discipline. The specialist evidence, mechanisms, ethics and authority remain domain-specific.
11. Common ways models fail
| Failure | What happened | Repair |
|---|---|---|
| Wrong target | The model predicts a proxy instead of the real decision target. | Restate the job and receiver. |
| Hidden assumptions | Users cannot see what must remain true. | Declare assumptions and operating range. |
| Overfitting | The model memorises existing data patterns. | Use held-out data and external tests. |
| Range failure | The model is used outside the conditions it learned or was validated on. | Restrict use or collect new evidence. |
| Proxy drift | The measured indicator stops representing the intended construct. | Revalidate the measurement relationship. |
| Model monoculture | One representation becomes unquestioned. | Test plausible alternatives. |
| False precision | Outputs show more digits than the evidence warrants. | Report uncertainty and practical resolution. |
| Authority leakage | A technical prediction is treated as permission to act. | Separate evidence from authority and decision rights. |
| No world return | Outputs are never checked against outcomes. | Measure what happened and update. |
12. How to read any model
- Purpose: What job is this model supposed to do?
- Target: What real system or process does it represent?
- Boundary: What is included and excluded?
- Variables: What features are represented?
- Assumptions: What must be approximately true?
- Inputs: Where did the data or starting conditions come from?
- Validation: What independent observations have tested it?
- Error: Where and how does it fail?
- Range: For which places, populations, scales and times is it justified?
- Alternative: What other model could explain the same observations?
- Decision: What action, if any, is the model actually authorised to inform?
- Return: What new observation would make us revise it?
13. Hostile test: the perfect-looking model
Imagine a model with beautiful graphics, tiny reported error and impressive historical performance. Now ask three questions:
- Was the error measured on data the model had already seen?
- Would a different plausible model make the same decision?
- What happened when the model met a new population, time period or operating condition?
If those answers are missing, the visual polish is not evidence of general validity. The model may still be useful, but its confidence must remain bounded.
14. Where Models fits in the wider How Things Work map
Models sits between representation and evidence. Useful neighbouring mechanisms include How Maps Work, How Measurement Works, How Evidence Works, How Classification Works, How Forecasting Works, How Optimisation Works and How AI Works.
These links are not synonyms. A map is a spatial representation; measurement produces observations tied to a reference or procedure; forecasting estimates future states; optimisation selects among feasible alternatives under an objective; AI may contain or use models. Keeping the ownership boundaries visible is what lets the concepts connect without collapsing into one another.
15. What this article does not claim
- It does not claim that one modelling method is best across all domains.
- It does not treat correlation as causation.
- It does not treat model fit as proof of truth.
- It does not treat AI output as an independent observation of the world.
- It does not convert population-level models into individual medical, legal, educational or financial decisions.
- It does not imply that technical capability creates authority to act.
16. Observable mastery test
You understand how models work when you can take an unfamiliar model and identify its purpose, target, boundary, assumptions, variables, inputs, outputs, validation evidence, uncertainty, operating range, plausible alternative and correction trigger—and when you can explain why a model may be useful without being literally identical to reality.
Authoritative source corridor
- NIST Research Data Framework — model, verification and validation concepts.
- NIST/SEMATECH Engineering Statistics Handbook: model building — selection, fitting and validation as iterative steps.
- NIST: Modeling Methodologies and Simulation for Dynamical Systems.
- National Academies: Ground Water Models—Scientific and Regulatory Applications — model limitations, verification, validation and the risk of false precision.
- National Academies: How Modeling Can Inform Strategies to Improve Population Health — models as purposeful simplifications whose causal assumptions require testing.
Governing idea: The model earns trust by remaining answerable to the world it represents.