What part of reality is being represented?
Define the system boundary.
Every model begins by deciding what counts as inside, outside or irrelevant.
Side 26
Models &
Representation
A study of how reality is compressed into something usable. Models reveal structure by omitting detail; representation becomes dangerous when the omissions disappear from view.
world→selection→representation→reasoning→decision
06model forms
05assumption tests
06distortion modes
26Side
A model is useful because it leaves things out.
The first question is not whether a model is “true,” but what task it is meant to support and which details can be safely ignored for that task.
What is the model for?
Explain, predict, optimize, communicate?
The same system may need different models for different tasks.
What features survive compression?
Which quantities or categories matter?
Selection makes some relationships visible and others disappear.
How are variables related?
Equation, graph, rule, simulation?
The form of representation determines which questions become easy to ask.
What action will depend on it?
What happens if it is wrong?
Model risk increases when decisions are high-stakes, irreversible or difficult to monitor.
Model disciplineusefulness comes from selective omission; risk comes from forgetting what was omitted
Different representations expose different structures.
A diagram, equation and simulation can all represent the same system while making different relationships salient.
Show structure visually.
Useful for components, hierarchy, sequence and causal or spatial relationships.
Encode quantitative relationships.
Useful when variables and functional relationships can be specified mathematically.
Represent space selectively.
Projection, scale and symbology determine what geography becomes easy to see.
Represent by categories.
Classification makes similarity and difference manageable but can harden fuzzy boundaries.
Represent behavior over time.
Useful when interactions, randomness or feedback make analytical solutions difficult.
Represent sequence and meaning.
Narratives compress causality and chronology into an interpretable chain, but can overstate coherence.
Assumptions are the hidden architecture.
Every model relies on conditions that may be explicit, implicit or inherited from the modeling tradition itself.
Boundary
What is excluded?
External forces may be treated as fixed even when they interact with the modeled system.
Linearity
Are effects proportional?
Linear approximations are powerful locally but can fail where thresholds or saturation matter.
Stationarity
Do relationships persist over time?
A model fitted to one regime may fail after structural change.
Independence
Are components treated as separate?
Shared causes can invalidate models that assume independent errors or failures.
Agency
Do people react to the model?
Predictions, rankings and rules can change behavior, altering the system being modeled.
A model can fit yesterday and fail tomorrow.
Validation asks whether the model performs adequately on the task and conditions for which it will actually be used.
Can it reproduce known data?
Necessary for many models, but weak evidence of generalization by itself.
Does it survive unseen cases?
Holdout data tests whether learned structure extends beyond the examples used to fit the model.
Would it have worked historically?
Useful when carefully designed, but vulnerable to leakage and retrospective tuning.
What happens outside normal conditions?
Extreme but plausible scenarios reveal fragility hidden by average performance.
Which assumptions drive the answer?
Small changes in uncertain parameters can expose unstable conclusions.
Does it travel?
A model can work in one population, location or period and fail elsewhere.
Representation changes what becomes visible.
Distortion is not always a flaw; sometimes it is the price of simplification. The danger is mistaking the representation for the world.
Variation disappears.
Averages can conceal subgroups, tails and structural differences.
Geometry is preserved selectively.
Maps trade off area, shape, distance and direction depending on projection.
Measured dimensions dominate attention.
Unmeasured qualities can become invisible in management and evaluation.
An abstraction becomes a thing.
Constructs such as “risk score” or “market sentiment” can be treated as concrete entities rather than model-dependent summaries.
External effects vanish.
Costs pushed outside the model can appear not to exist.
Messy causality becomes a clean story.
Retrospective explanations often compress contingency into an apparently inevitable sequence.
Read the model before reading the output.
Ask what the representation had to assume before deciding whether its result is persuasive.
A dashboard says performance is improving.
Inspect which variables were selected, what was excluded, whether definitions changed and whether improvement in the metric corresponds to improvement in the underlying objective.
A simulation shows a 20% capacity increase solves congestion.
Check arrival assumptions, behavioral response, service-time distributions, bottleneck migration and whether the modeled system boundary excludes upstream or downstream constraints.
A map makes one region look dominant.
Check projection, area scaling, symbol size, normalization and whether absolute totals or per-capita values are being shown.
A simple model beats a complex one.
Complexity is not automatically better. If the simple model generalizes better, is more stable and supports the actual decision, added detail may be noise rather than information.
Models.Behaving.Badly.Emanuel Derman · model limits and misuse
The Model ThinkerScott Page · multiple-model reasoning
Visual ExplanationsEdward Tufte · representation and evidence
Models as MediatorsMorgan & Morrison · models in scientific practice