Review tails, drivers, scenarios, and readiness

Review the complete distribution before focusing on one policy percentile. Examine mean, spread, minimum, maximum, P10, P50, P80, P90 and P95, the histogram and cumulative curve, baseline-overrun probability, approved-boundary confidence, convergence, and warnings. In integrated mode, inspect cost–duration scatter and joint confidence rather than assuming both dimensions move together.

Driver rankings report association between modeled inputs and outturns. Rank correlation, occurrence frequency, and mean contribution answer different questions. A tornado or sensitivity row identifies leverage inside the encoded model; it does not establish causation, validate an estimate, or tell a manager which response is feasible. Check whether a dominant driver is real exposure, duplicate modeling, weak evidence, or simply a very wide assumption.

Scenario and readiness review

Compare inherent and residual cases only when responses are explicit and variable definitions remain comparable. Use the same seed and stable variable set where possible. State every exclusion, correlation setting, override, and changed assumption. Do not present a favorable residual case as mitigation benefit when its values are unsupported.

Validation separates valid for simulation from ready for decision. A run can be computationally valid yet lack evidence, owners, stable convergence, approved boundaries, or a recommendation. The accountable person records adopted contingency and schedule reserve, rationale, trade-offs, conditions, next steps, title, and date. The engine and AI must never write or approve that recommendation. Continue to reporting and audit trail. ## Interpretation discipline

Keep inputs, derived results, and human decisions visibly separate. Every material number needs linked evidence or an explicit assumption flag and basis. Unknown is not zero. A percentile is conditional on this model, scenario, seed, engine version, and source version; it is neither a guarantee nor an approval. Any material model edit makes an earlier result stale.