Construct a coherent decision tree

A good tree is a chronological model of choices and information, not an exhaustive map of everything that might happen. Begin at the root with the next decision in scope and move left to right through actions, uncertain events, later choices, and terminal consequences.

Three node types

1. Decision node — a choice available to the named decision authority. Its branches are alternatives, not random outcomes, so they do not receive probabilities.
2. Chance node — an uncertainty not under direct control. Its outgoing outcomes must be mutually exclusive and collectively exhaustive at that point, and their conditional probabilities must sum to 1 within tolerance.
3. Terminal node — the end of a complete path inside the chosen horizon. It has no outgoing branches and records the consequence value and its basis.

Branches may also carry incremental costs or benefits. State clearly whether a number belongs on a branch or at a terminal node and count it only once. A terminal value should describe the complete remaining consequence in the same currency, utility, cost, or score unit as every other terminal value. Never add money to an ordinal score as though the units were interchangeable.

Construction sequence

• Define the root and add distinct, feasible alternatives, including defer or status quo when material.
• After each alternative, ask what important uncertainty becomes known next. Split only when the outcome could change consequences or a later choice.
• Add later decision nodes only when that choice will genuinely remain available at that time.
• Continue until each path ends in a terminal consequence within the time horizon.
• Link each probability and value to a basis: source, calculation, expert elicitation, or explicit assumption.
• Validate cycles, orphan nodes, duplicate destinations, missing labels, incomplete chances, missing terminal values, and unreachable content.

Probability discipline

Every chance probability is conditional on the path already taken. “Permit approved” after a staged design may require a different probability from approval after the full design. Record the conditioning context and date. Use relevant frequency data when available; otherwise document how expert judgments were elicited, reconciled, and calibrated. More decimal places do not create better knowledge.

Keep the tree decision-relevant

Branching grows rapidly. Combine outcomes only if they lead to equivalent consequences and subsequent choices; otherwise aggregation can hide risk. Conversely, do not split on uncertainty that cannot alter a choice or material consequence. A compact, reviewable tree with transparent limitations is more useful than a large diagram that nobody can audit.

Visual position is presentation metadata. Node IDs, branch relationships, values, evidence, and warnings are the authoritative model. Moving a node must not alter the analysis.