Models, Realism, and Useful Fiction

By the end of this lesson, you will be able to...

• Separate a model's predictive, explanatory, ontological, and intervention claims.

• State which parts of a model are measured, idealized, or merely convenient.

• Use a model productively without treating its successful predictions as proof that every internal entity is literally real.

Idea in one sentence: A scientific model is a selective construction that can be useful, explanatory, or intervention-guiding without being a complete mirror of reality.

Core Insight

Imagine the support team builds a model that assigns a probability that an answer contains an unsupported claim. It predicts well on the evaluation set, so someone says, “The model understands factuality.” That sentence quietly combines four different claims: it predicts labels, it explains why answers fail, it represents a real property called factuality, and it can decide which intervention to deploy.

Those claims may have different evidence. A model can forecast accurately while using internal features that are not humanly meaningful. A simplified model can explain a mechanism while ignoring details needed for a production decision. A useful abstraction can guide an intervention even when no literal object in the world corresponds to every symbol.

The previous lesson showed that theories organize anomalies and standards. This lesson asks a narrower question: What kind of contact with reality does a model claim? The answer should be specific enough to guide testing and modest enough to survive the model's boundary conditions.

The central trade-off is that useful models drive progress, but success can tempt overinterpretation. Simplification makes reasoning possible; it also creates a path from “works here” to “is literally true everywhere.”

What a Model Selects

A model is a representation built for a purpose. It selects variables, relations, scales, and outputs while suppressing other details. That selectivity is not automatically a flaw. A subway map is useful because it preserves connectivity and discards geographic shape. A frictionless plane is useful because it isolates a relationship before real surfaces are added. A queueing model can estimate waiting time without simulating every CPU instruction.

When evaluating a model, ask four questions:

1. What target is being represented? A measured outcome, a mechanism, a system structure, or a decision threshold?
2. Which features are preserved? Correlations, causal direction, conservation, ordering, or response to intervention?
3. Which features are deliberately omitted? Noise, heterogeneity, history, feedback, or measurement error?
4. For what use is the model adequate? Forecasting, explanation, diagnosis, control, or communication?

The same model can be adequate for one use and misleading for another. A traffic model may forecast average flow while failing to explain an individual driver's route. A language model may rank likely answers while lacking a reliable account of why a source entails a claim.

Four Claims That Must Be Separated

Prediction

The model maps inputs to expected observations. A retrieval classifier that predicts evaluator labels is successful if its forecasts are calibrated and useful on specified data. Predictive success is empirical and local: it depends on the population, metric, comparison baseline, and drift.

Prediction does not require that the model's internal features correspond to real entities. A weather model can predict tomorrow's temperature with approximations that are not literal miniature weather systems.

Explanation

An explanatory model connects observations through a mechanism or dependency. It says why a change should occur and what else should follow. For the support assistant, “low source coverage causes unsupported claims because retrieval cannot supply an entailing passage” is an explanatory hypothesis.

Explanation is stronger than curve fitting, but it is still constrained by the model's abstractions. A causal story built on an unmeasured proxy can sound mechanistic while hiding a confounder from lessons 003–004.

Ontology

An ontological claim says what the model's entities or structures are like in the world. Does “retrieval relevance” name a stable property, or is it a score produced by one index and embedding model? Does a latent “risk factor” correspond to a real mechanism, or is it a useful summary of correlated observations?

Ontology is not settled by prediction alone. Two models can predict equally well while making incompatible claims about what exists. Additional interventions, independent measurements, and cross-context constraints are needed.

Intervention

An intervention claim says that changing a model variable will change a target outcome. A model that predicts which sessions contain unsupported claims need not tell us that increasing its “confidence” will improve answers. For intervention, the variable must connect to a manipulable process, and the model must survive the causal checks from lesson 004.

Keep these claims in separate boxes. A model can be predictive but not explanatory, explanatory but not intervention-ready, or ontologically cautious while still guiding a practical decision.

The Retrieval Model as a Case Study

Suppose the team uses a simple scoring model:

support_score = 0.6 * source_entailment
              + 0.3 * retrieval_relevance
              - 0.1 * answer_uncertainty

The score is not a substance inside the assistant. It is a constructed quantity whose weights, labels, and features were chosen for a task. It may still be useful.

If the score predicts labels on a held-out set, that supports a predictive claim. If controlled tests show that changing source coverage changes the score and the claim outcome as expected, the explanatory and intervention claims gain support. If the score fails under a new domain, the model's scope or measurement must be revised; the failure does not prove that “support” is unreal.

Idealization and Useful Fiction

An idealization intentionally violates some facts to reveal a relationship. Physicists use point masses, perfectly elastic collisions, or frictionless surfaces. Economists may model agents as optimizing under simplified information. Engineers may treat a dependency as instantaneous or a component as independent for a first approximation.

Calling such a model a “fiction” can be helpful if it means “constructed representation,” not “arbitrary lie.” A useful fiction earns its place by:

• isolating a mechanism;
• making a risky prediction;
• explaining a pattern across more than one case;
• guiding measurement or intervention;
• and stating where the idealization breaks.

The ideal gas law is not literally true for every gas at every pressure, yet it captures a stable relation in a defined regime. Its usefulness depends on knowing the regime and the omitted interactions. The same discipline applies to a model of support quality: a scalar score may be adequate for triage while inadequate for adjudicating a safety-critical answer.

Realism Positions Without a False Choice

Three positions help organize discussion, but they are not mutually exclusive slogans.

• Instrumentalism: Judge a model primarily by what it predicts and helps us do. Be cautious about claims that its entities are real.
• Scientific realism: Successful, robust, and independently constrained models can give us reason to believe that at least some of their structures or entities track reality.
• Structural or selective realism: We may have stronger grounds for believing that stable relations and structures are captured than for believing every unobservable entity or parameter is literally represented.

In practice, a scientist can be instrumentalist about a temporary forecasting model and realist about a well-tested mechanism. The useful question is not “Which label do I wear?” but “Which part of this model has earned which level of commitment?”

A good commitment ledger looks like this:

This ledger prevents a successful prediction from silently upgrading every other claim.

Model Comparison and Model Pluralism

Different models can be useful for the same system. A queueing model estimates latency; a causal graph identifies bottlenecks; a simulation explores feedback; a dashboard summarizes operations. They may disagree because they answer different questions or operate at different scales.

Prefer model pluralism over one-model worship. Compare models by:

• target and purpose;
• predictive performance under relevant shifts;
• explanatory mechanism;
• intervention validity;
• sensitivity to omitted variables;
• interpretability and auditability;
• and cost of data, computation, and false decisions.

If two models make the same forecast but recommend different interventions, the disagreement is valuable. It identifies an untested causal or ontological assumption. Design a measurement or experiment that separates them.

Boundary Conditions and Failure Modes

Reification: Treating a score, category, or latent variable as a thing that exists independently of the measurement process. Keep the operational definition visible.

Overfitting the explanation: A model can fit historical data because it memorized quirks. Test on fresh populations, interventions, and adversarial cases.

Scale confusion: A model at aggregate scale may not describe individuals. Average effects can conceal heterogeneous mechanisms.

Purpose drift: A model built for prediction gets used for causal control or moral judgment without new evidence. Re-open the claim ledger when the decision changes.

False precision: Parameters with many decimal places can still encode uncertain constructs or unstable estimates. Match precision to measurement quality.

Model monopoly: A single representation becomes the only permitted way to ask questions. Invite alternative models, especially from domains or stakeholders the current abstraction omits.

Active Checks

Check 1: Prediction without intervention

A classifier predicts which support answers will receive an “unsupported” label with 95% accuracy. Can the team conclude that increasing the classifier's confidence will reduce unsupported claims?

Answer: No. The result supports prediction, not intervention. The classifier may be a useful monitor, but changing its confidence does not necessarily change the answer-generating process.

Check 2: Useful idealization

A frictionless queue model predicts average waiting time well during normal load but fails during overload because feedback and retries dominate. Is the model useless?

Answer: No. It has a defined regime of adequacy and a visible boundary. Add a feedback model or switch representations when the decision concerns overload behavior.

Transfer Practice

Choose a model you use at work or study. Fill out this four-part card:

1. Prediction: What observations does it forecast, and on what population?
2. Explanation: What mechanism does it claim connects the variables?
3. Ontology: Which entities or structures are you willing to treat as more than convenient labels?
4. Intervention: What could you change, and what experiment would test whether the model's recommendation works?

Add the model's idealization and failure boundary. Then write one sentence of calibrated commitment: “I trust this model for ___ because ; I will not infer ___ until .” This prepares the capstone, where you will evaluate a full scientific claim from hypothesis through decision implication without confusing confidence with certainty.

Resources

• [ARTICLE] Scientific Realism - Stanford Encyclopedia of Philosophy - Focus: what successful theories justify us in believing.
• [ARTICLE] Models in Science - Stanford Encyclopedia of Philosophy - Focus: representation, idealization, and model pluralism.
• [ARTICLE] Scientific Method - Stanford Encyclopedia of Philosophy - Focus: prediction, explanation, and testing.
• [BOOK] The Structure of Scientific Revolutions - Focus: how frameworks shape model-building and problem choice.

Key Takeaways

• A model's predictive success does not automatically establish its explanation, ontology, or intervention advice.
• Idealizations are useful when they isolate a mechanism and state their regime of validity; they become dangerous when their omissions disappear from view.
• Commit selectively: demand stronger evidence for literal or causal claims than for a bounded forecasting tool.
• Model pluralism and explicit failure boundaries make scientific reasoning more robust than treating one successful model as a complete picture of reality.