Decision Tree & Learning Value

Use this workbench when a decision contains sequential choices, uncertain events, materially different consequences, or a real option to learn before committing. It turns the reasoning into an inspectable tree and uses rollback to compare complete strategies under one declared value model.

It does not predict which outcome will occur. A probability is a quantified state of knowledge supported by data, expert judgment, or an explicit assumption—not a fact about the future. Expected value is a probability-weighted decision metric, not the outcome to expect on a particular project. Rollback identifies the preferred policy given the model, objective, constraints, and inputs; it does not certify that the policy is safe, fair, lawful, fundable, or optimal in the real world.

What a complete model contains

• a precise decision question, owner, objective, scope, horizon, stakeholders, constraints, assumptions, and value perspective;
• decision, chance, and terminal nodes connected without cycles or unreachable fragments;
• feasible alternatives and conditional outcomes that do not hide a meaningful “do nothing,” defer, redesign, or stop option;
• probabilities whose source, confidence, range, as-of date, and reviewer are recorded in the probability basis or linked evidence notes, with explicit assumption status;
• branch increments and terminal consequences in one unit, with costs and benefits counted once;
• rollback results, chosen policy, probability-weighted risk profile, and unresolved validation warnings;
• EVPI, EVSI, information cost, and net learning value only where the required information structure is valid;
• built-in one-way chance-probability sensitivity with sampled approximate switch intervals, plus limitations, reservations, and a version-bound human review.

Start with the decision—not the diagram

Write the decision as a choice the named authority can actually make by a relevant date. Identify what is inside the model, which consequences matter, whose perspective defines value, and what would make an option infeasible. For a highway project, for example, separate the decision to proceed or stage construction from uncertain geotechnical conditions, then include schedule, safety, environmental, community, and distributional consequences that matter to the decision. Do not collapse every consequence into money when monetization would be misleading or unauthorized.

Appropriate and inappropriate uses

Decision trees are useful for a bounded sequence with a manageable number of consequential branches. Use scenario planning for broad, interacting futures; simulation for many continuous uncertainties; and a multi-criteria method when important values cannot be coherently reduced to a single objective. A tree can complement these methods.

NASA describes decision analysis as support for authorities choosing with incomplete knowledge and stresses documenting assumptions and limitations. The UK Green Book notes that decision trees can clarify sequential or difficult-to-reverse choices, while warning that probabilistic precision can be spurious. Match the rigor to the stakes and always preserve accountable human judgment.

Human and AI roles

The AI Assistant may help draft a frame, identify missing branches, explain formulas, or propose sensitivity questions from supplied context. It must not invent probabilities, values, test accuracy, evidence, stakeholder agreement, legal compliance, or expert consensus. Suggestions remain proposals until an authorized person reviews and applies them.