Scientific, Medical, and Social Worlds › Scientific, Medical, and Social Worlds

A short case history can fit two models and still leave their answers to a new intervention far apart. In the synthetic experiment below, we observe weeks 0 through 24 of rising incidence, then reduce the simulator's contact multiplier by 70 percent at week 25. A fitted compartment model responds through its transmission term. An autoregression without a contact input continues its fitted trend. The compartment model also makes an inaccurate forecast: its recovery parameter is weakly constrained by the selected observations.

The disagreement has two sources: different transition assumptions, and uncertain parameters within the compartment model. A good fit to this history does not establish either model's intervention response. The worked example makes that distinction visible without treating a synthetic outbreak as evidence about a real disease.

We will compare fits to the same observations, not claim a held-out forecasting contest or clinically plausible parameter estimates. Within the compartment model, refitting transmission and initial prevalence at several fixed recovery rates leaves the training loss nearly unchanged while changing predicted post-intervention totals. This shallow loss profile illustrates practical uncertainty under the specified data and simulator. It is not proof of exact observational equivalence.

The previous chapter, Digital and Software Environments, asked whether a proposed action actually changed the intended software state. Here there is another question: does the model represent how that change affects later outcomes? We retain $s_t$ for the underlying state, $o_t$ for an observation, $b_t=p(s_t\mid o_{\leq t},a_{<t})$ for a belief about hidden state, and $a_t$ for an action. An action can be a dose, a contact multiplier, or a specified external forcing. None of these mathematical action variables automatically describes an implementable or causally identified real intervention.

Scientific and medical studies can collect experimental evidence. Social experiments can randomize participants or clusters. What is unavailable is both mutually exclusive outcomes for the same unit in the same episode; an experiment instead compares suitably assigned units. Other questions concern interventions that are unsafe, expensive, or impossible to randomize, such as changing a common national policy regime. Robotics and driving also have unsafe or unsupported actions. The relevant distinction is the evidence available for the particular query, not a categorical division between experimental engineering and observational science.

The latent-state tools from Part V and the planning tools from Part VII provide a vocabulary for these problems. Whether they support an inference depends on the observation process, the action semantics, and the validation regime. This chapter develops that vocabulary through molecular and materials models, weather and climate, physiological dynamics, and economic and social simulation. The examples are educational, not clinical or policy recommendations.

Four questions help separate kinds of model quality:

• Predictive fidelity: does the model reproduce held-out observations under the stated evaluation distribution and metric?
• Representation quality: does its state retain the information needed for the task, such as delayed physiological effects rather than only the latest measurement?
• Uncertainty quality: do probabilities or intervals have their claimed calibration under a specified regime? Calibration on one regime does not establish calibration after a shift.
• Decision usefulness: does the model support better choices under the feasible actions, costs, constraints, and losses of the decision problem?

These questions can interact but are distinct. A treatment-effect error can change a preferred action; it need not do so if the action ranking remains unchanged. An emulator can fit historical observations and still give unjustified confidence under an untested forcing. Conversely, confidence alone is not evidence that a forecast is wrong. Each evaluation needs its own evidence.