Validation, human review, AI, and reporting

Before decision use, resolve high-severity findings and confirm: portfolio authority and boundary; comparable objectives and units; complete capacity, demand, and score cells; field-level provenance; mandatory work and category rules; dependencies, exclusions, and synergies; scenario logic; solver quality; displaced work; bottlenecks; sensitivity; affected stakeholders; safety, environmental, equity, legal, procurement, engineering, financial, and operational reviews.

The human recommendation is separate from the engine result. Record the chosen scenario or portfolio, rationale, decision conditions, next actions, reservations, reviewer identity and time through the governed workflow. Material changes to inputs or constraints make prior analysis and review stale. Preserve the source fingerprint, version, validation findings, and exact-versus-heuristic status.

readyForDecision is necessary but not sufficient for publication: the active scenario must also recompute as feasible with status optimal, heuristic, or evaluated on the exact active version. Persisted data is never silently truncated. The separate compute envelope is 1,000 initiatives, 5,000 relationships, 50 periods, 100 resources, 100 objectives, 100 scenarios, and 200,000 modeling entries. Modeling entries are capacity cells, demand cells, objective-score cells, scenario capacity overrides, and locked-in, locked-out, or manual-selection references. Exact search is additionally bounded to 24 optional candidates and 250,000 visited nodes; exceeding that envelope requires a disclosed heuristic result, not a false optimal claim.

AI role

The assistant may summarize supplied evidence, draft editable inputs, identify gaps, explain deterministic outputs, and propose questions. It must use exact authorized evidence identifiers, mark unsupported material numbers as assumptions, and keep every proposal in Review/apply. It must not invent evidence or numbers, calculate or persist derived results, change the active scenario or recommendation, attest a reviewer, claim global optimality without engine proof, or say a proposal was applied.

Report content

Include frame and scope; model/version and source fingerprint; objectives and weights; candidates; capacity and demand ledger with provenance; interactions and rules; scenarios; validation; selected modeled set; solver quality and limits; utilization and bottlenecks; category balance; frontier and sensitivity when entitled; evidence and assumptions; human recommendation and conditions; unresolved issues; AI disclosure; and methodological limitations. The PDF is a decision-support record, not a certificate or authorization.

Return to the method overview.