Optimization
For Both Executive and Technical ReadersInference tells you what is true. Inquiry tells you what to learn next. Optimization searches over everything you could observe or do, and returns the combination that best meets your objective under your constraints.
01 The Gap
Inference tells you what is true. It doesn’t tell you what to do.
Inference and information value both stop at knowledge. A system that can tell you what is true and what to measure next, but cannot tell you what to do about it, hands the most consequential step back to human intuition, unsupported, unjustified, and unauditable.
Optimization crosses into action. It is the layer that turns an explicit model into a recommended course of action, with the objective stated, the constraints respected, and the rejected alternatives recorded. The constraints are the point: “maximise survival” is a goal; “maximise survival for under a thousand dollars” is a decision.
02 The component
Searches over actions and returns the best one, constrained and justified.
Optimization is a Cognitive Primitives, Action component. It searches over combinations of observations and, in causal models, of actions, against an explicit objective under explicit constraints, and returns an OptimizationResult: the best feasible bundle, the objective value achieved, and what was rejected and why.
| Field | What it holds |
|---|---|
| Objective & constraint | Stated in plain terms, what's being maximized, and what boundary it can't cross |
| Recommended | The action set chosen, its cost, and the outcome distribution it produces |
| Rejected | Every alternative considered and turned down, each with the specific reason, not silently discarded |
The rejected field is required. A recommendation without a record of what was considered and why it was not chosen is not auditable. The optimizer must show its work.
03 How it works
Three levels: observe, intervene, constrain.
Three levels of reasoning, in order of ambition:
Inference. What is true? Reads the current posterior. Rung 1.
Value of Information. What should I learn next? Ranks candidate observations by expected decision impact. Rung 1 over the model structure.
Evidence Optimization. What actions best achieve the objective, under constraints? Searches over combinations of observations and interventions. Rung 2 and Rung 3, do() queries in a constrained search.
Optimization generalises Inquiry in two directions: it searches combinations (not just single probes), and it includes actions (not just observations). The search respects the model’s causal structure, it cannot recommend an action that is not represented as an intervention node.
04 The Brains
The model deciding, not predicting.
Optimization is the brains deciding what to do, constrained, justified, and auditable.
There is no .bayes file for Optimization, it operates on any domain-specific domain model. The OptimizationResult is appended to the AuditRecord. The objective and constraints are stated inputs, not hidden assumptions.
| Field | Value |
|---|---|
| Objective & constraint | Maximize P(Survival), total cost ≤ $1,000 |
| Recommended | EarlyAntibiotics ($200) + Lactate test ($50), P(Survival) = 0.87 |
| Rejected | ICU Admission, exceeds the budget constraint |
05 Query in plain English
What should we do?
Optimization answers the question after the posterior: what should we do?
The audit trail is the OptimizationResult, including rejected alternatives.
06 Where it sits
Downstream of Scenarios. The decision layer of the Action cluster.
Optimization is the decision layer of the Cognitive Primitives, Action cluster. It sits downstream of Scenarios (09, which maps the landscape) and upstream of the governance handoff (12 BPMN Integration, which receives the recommendation as a certified artifact).
Upstream: the current posterior (from Execute), the ScenarioComparison (from 09 Scenarios), and the stated objective and constraints from the analyst. Downstream: the OptimizationResult is appended to the AuditRecord and consumed by 12 BPMN Integration. It interacts closely with 05 Inquiry (which VoI calculation it extends) and 11 Monitoring (which tracks whether the recommended action had the expected effect).