The operations vice president has thirty-one candidate interventions ranked by expected value. Each is specified at the operational level: increase SKU output by twenty percent, renegotiate a packaging contract, defer a line changeover. The list appears actionable. But the interventions live in a tightly coupled production system. Some need the same labor. Some are mutually exclusive—you cannot surge one SKU while reallocating that same labor elsewhere. Some have prerequisites: the contract renegotiation must complete before the packaging change begins. The total cost of all thirty-one exceeds the operations budget. The operational complexity of changing thirty-one things simultaneously would create more disruption than any intervention is designed to address. The ranked list cannot be executed as a list. This is the central paradox that this chapter resolves.

Chapters 14 through 17 established a Living Model—a causal graph validated through sensitivity analysis, backdoor criterion application, instrumental variable methods, and structural validation. That model is scientifically sound. But a validated graph does not tell an operations team which interventions to execute when labor, budget, and calendar are finite and interventions interact. The board has asked for a recommendation by Monday. The VP needs not a ranked list but a coherent portfolio: roughly six interventions that fit the budget, do not conflict, and can be sequenced through the quarter. This chapter bridges validation and decision. It shows how parameterization transforms the graph into a generative model, how counterfactual reasoning ranks individual interventions, and how portfolio optimization converts a ranked list into a feasible set of actions under constraint.

Three kinds of resource constraints bind the decision. Labor is the most rigid: the company runs eleven production lines across three plants. Each line has fixed staffing; changing output on one line requires reallocating labor from another. That reallocation is the intervention itself. A decision to surge SKU 4421 by twenty percent means pulling labor from line two to execute it. If another intervention also needs line-two labor, they are mutually exclusive. Budget is the second constraint: even if labor could be magically stretched, the operational budget cannot absorb all thirty-one interventions at once. Calendar is the third: some interventions have prerequisites, and the quarter has fixed length. Atop these sit interaction effects. Some interventions are synergistic: together they amplify each other's impact. Others are antagonistic: executing both creates disruption that cancels their individual gains. The portfolio must navigate all four constraints simultaneously.

Parameterization is the bridge from Chapter 17's validated causal graph to the counterfactual predictions needed for decision-making. The graph specifies structure: demand depends on price and marketing, output depends on labor and machinery, profit depends on revenue and cost. Structure alone does not generate numbers. Parameterization asks: how strongly does demand respond to price? If we increase price by ten percent, does demand fall by five percent or fifteen percent? The answer is encoded in path coefficients, elasticities, and functional relationships. For each variable in the graph, we write its value as a function of its parents plus noise. In the simplest case—linear parameterization—that function is a weighted sum. The weights are the path coefficients, estimated from observational data using regression methods with deconfounding sets from the backdoor criterion. The result is a fully specified model: given values for the exogenous variables, the model generates a prediction for every endogenous variable.

The simplest parameterization is linear. Each variable is written as a weighted sum of its parents plus noise. The weights—the path coefficients—are estimated using standard regression with deconfounding sets identified through the backdoor criterion discussed in Chapter 8. When the true relationship is linear, this method produces unbiased estimates. The advantage is simplicity: the model is easy to fit, easy to inspect, and easy to verify. You can look at the coefficients and check whether their magnitudes make intuitive sense. You can examine residual plots for violations of the linearity assumption. You can compare predictions to held-out data. But many real systems are nonlinear, and in exactly the regions where operational decisions matter most. The relationships between price and demand, marketing spend and acquisition, or staffing and output rarely maintain linear form across wide ranges. A linear model fitted to data where prices range from one to five dollars will fail badly when asked to predict at ten dollars.

Real mechanisms are frequently nonlinear. Price and demand follow a curve, not a line: increasing price from one dollar to two dollars may reduce demand far more than increasing it from nine dollars to ten dollars. Marketing spend shows diminishing returns; the first thousand dollars in acquisition spend produces more customers than the thousandth thousand dollars. Staffing and throughput exhibit saturation: adding workers to a production line increases output, but as the line fills up, the benefit of each additional worker falls. A linear model fitted to data near the current operating point will capture these nonlinearities poorly. It will estimate an average slope across the data. But when the recommendation calls for surging output by twenty or thirty percent—pushing the system away from its current equilibrium—the average slope is not the local slope. The model predicts badly. A mature operations system needs parameterizations that capture nonlinearity in the regions where decisions will operate.

When linearity fails, parameterization options expand. Generalized linear models extend the linear case to handle outcomes that are bounded, binary, or count-valued: a logit model for binary choices, a Poisson model for counts, a beta model for quantities constrained between zero and one. These are parametric families: you specify the functional form, estimate the parameters, and diagnose fit. Spline-based methods offer more flexibility: they fit piecewise polynomials that can bend and twist without forcing a specific functional form. Modern approaches use deep structural causal models: neural networks learn the mapping from parents to child, preserving the graph's causal structure while allowing arbitrary nonlinearity. The trade-off is cost. As the function family grows richer, estimation demands more data, and diagnostic verification becomes harder. A linear model is easy to validate. A deep model often is not. A mature deployment typically uses linear parameterizations for most variables and richer ones where domain science demands it—where you know the mechanism is nonlinear and you have data to support estimation.

The choice of parameterization depends on the region of decision space the recommendation will occupy. If the recommendation asks for a ten percent increase in output—a small change near the current operating point—linear parameterization is usually sufficient. The relationship between input and output looks linear over a small range, even if it is nonlinear globally. The linear model is a local approximation valid in a neighborhood around the current equilibrium. But if the recommendation calls for a two-hundred percent increase—a large change that pushes the system far from equilibrium and into a new operating regime—the linear approximation breaks down. The true nonlinear curve diverges from the linear model's prediction. The recommendation becomes wrong. A VP's decision is often far-field: the board does not ask for incremental improvement near the current state; it asks for changes large enough to matter strategically. This pushes the system away from equilibrium, and away from the validity of linear approximations. Mature systems use linear parameterizations for safe, small changes and richer families for the aggressive decisions that define strategy.

Chapter 6 introduced the autonomy-of-causal-mechanisms principle: the fact that a causal edge exists does not imply that changing it affects other edges. The mechanism by which price influences demand is independent of the mechanism by which labor influences output. This principle is not merely philosophical. It is an implementation requirement that carries through the full architecture. When the model is built and deployed, each parameterized causal relationship should be a self-contained, replaceable code module. If you discover that the price-elasticity relationship is nonlinear and needs a spline fit instead of linear regression, you should be able to swap the module without touching the modules for demand-aggregation, cost-calculation, or profit-function. This modularity enables iteration: as domain science improves and data grows, you can upgrade individual mechanisms without rebuilding the entire model. It also enables debugging: when a prediction is wrong, you can test each mechanism independently to isolate the source of error. A monolithic, tightly-coupled model where every mechanism depends on implementation details of every other mechanism defeats this capability.

The chapter title announces the destination: from graph to decision. The path has four steps. First, parameterization: annotate the graph with coefficients, functional forms, and parameter estimates, turning a structural claim into a generative model. Second, counterfactual reasoning: given a proposed intervention—increase SKU 4421 output by twenty percent—use the parameterized graph to simulate the outcome. What happens to labor allocation, to cost, to profit? What is the expected change in the company's primary outcome, and what is the confidence interval around it? Third, ranking: compute expected value for all thirty-one candidate interventions; order them by the magnitude and certainty of their impact. Fourth, portfolio optimization: recognize that interventions interact, that resources are finite, and that some are mutually exclusive; use constrained optimization to select the subset that maximizes total expected value subject to budget, labor, and calendar constraints. This chapter unpacks the first step in depth: how parameterization works, what functional forms to choose, and the architectural principles that keep models maintainable.

Chapters 14 through 17 built and validated a causal graph through sensitivity analysis, backdoor application, instrumental-variable reasoning, and structural tests. The graph is causally sound. But causality is qualitative: A causes B, C causes B, B causes D. These claims are true or false, but they do not tell you how much A causes B, or whether an intervention on A will produce a visible change in D. Parameterization supplies the magnitude. By estimating path coefficients, elasticities, and functional relationships, it converts the graph from a qualitative causal claim into a quantitative predictive engine. Given values for exogenous variables and proposed interventions, the parameterized model generates point estimates and confidence intervals for every outcome of interest. This moves the analysis from validation to decision. The operations VP is not asking: Does labor affect output? She knows it does. She is asking: If we pull two workers from line two, how much does output on line two fall, and how does that interact with our plan to surge SKU 4421 using labor from lines two and three? Only a parameterized model answers that question.