Use the AI assistant as a reviewable modeling partner

The assistant can explain method selection, ask focused setup questions, draft a frame, organize supplied estimate items or activities, turn source statements into candidate risk events, flag possible double counting, and propose scenarios or dependency questions. It may interpret an exact server-generated snapshot of validation, percentiles, convergence, confidence, and drivers when the source version and entitlements match.

Every change appears under the dedicated proposal contract and remains pending until an authorized person reviews and applies it to the same active editable revision. The preview must show sanitized content, evidence linkage, explicit assumptions, normalization adjustments, and rejected fields. Stable IDs and normal model APIs preserve relationships and conflict checks.

Non-negotiable guardrails

The assistant must not invent a probability, estimate, duration, distribution, coefficient, response effectiveness, evidence item, citation, or expert consensus. It leaves an unknown null or marks an unsupported material number as an explicit assumption with a basis. It does not convert qualitative words into precise numbers without a declared human scale, silently assume independence, repair a non-PSD matrix, repair a cyclic network, or infer residual exposure from inherent exposure.

The assistant never computes Monte Carlo results itself, claims a run occurred without attached server output, persists derived results, selects the active scenario, or writes the human recommendation. It describes P80/P90 as conditional modeled confidence—not certainty, guarantee, approved reserve, or recommended funding. It treats sensitivity as association, not causation. It cannot auto-apply an open-model proposal, change status or locks, bypass limits, or expose private source content. Return to the method overview. ## 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.