Candidates, evidence, and uncertainty
Use comparable definitions for every candidate: scope, owner, category, lifecycle status, expected benefit, estimated cost, value score, risk score, confidence, objective scores, and resource demand. Separate “candidate,” “committed,” “deferred,” and “excluded.” A mandatory or committed flag is a hard modeling choice and requires an accountable rationale; it must not be inferred from enthusiastic language.
Expected benefit and estimated cost are reported context; neither is a hidden solver term. estimatedCost does not constrain selection by itself. If money is a hard limit, define a monetary resource in one consistent unit, record its capacity by period, and record each candidate’s demand in every applicable period. expectedBenefit influences selection only through an explicit valueScore or objective score. Keep the raw benefit visible, but do not double count it across score channels.
Each material estimate needs field-level provenance. Link an evidence item by its exact authorized identifier and record the estimation basis. If a number comes from a user scenario or expert judgment, mark it as an assumption and state how to validate it. Duplicate titles are not safe identifiers, a citation does not automatically prove every linked claim, and AI-generated prose is not evidence.
Represent uncertainty honestly. Confidence is not probability, risk is not simply cost, and a point estimate should not hide an important range or dependency. Use separate scenarios or sensitivity to test credible alternatives. Do not let missing evidence make a candidate look artificially inexpensive or unconstrained.