Review, reporting, and safeguards

A professional report lets another qualified person reproduce the logic and challenge the judgment. It is not a certificate, approval, forecast, or substitute for the underlying model. Export only the normalized model version that was actually reviewed, and label stale calculations or reviews after any substantive edit.

Readiness review

Before using a recommendation, confirm that:

• the owner, authority, decision date, scope, objective, horizon, value perspective, stakeholders, constraints, and mandatory requirements are explicit;
• all reachable paths are structurally valid and each chance node is complete;
• probabilities and values have units, evidence or assumption bases, dates, and responsible reviewers;
• feasible alternatives and meaningful stop, defer, stage, or redesign options were considered;
• rollback follows the optimization objective and preserves the full conditional policy;
• the risk distribution, unacceptable threshold, built-in one-way probability sensitivity and its sampled approximate switch intervals, and material non-monetized consequences have been reviewed;
• EVPI or EVSI labels match the calculation actually performed, and all information costs and delays are visible;
• affected perspectives, adverse and distributional effects, accessibility, privacy, safety, environmental duties, and legal or procurement constraints have not been hidden in an average;
• unresolved high-severity warnings, dissent, reservations, limitations, and evidence gaps are prominent.

Report composition

The report should contain: decision frame; model/version identity and timestamps; value model; tree inventory and readable path or policy summary; probability and consequence bases; rollback and recommended policy; risk profile; learning opportunities and information-value results; built-in one-way probability sensitivity and sampled approximate switch intervals; any external follow-up analysis clearly labeled as such; evidence register; warnings and limitations; human review status; and methodology notes.

Show units on every value, distinguish inputs from calculations, and use “not calculated” rather than zero when analysis is invalid or unavailable. A PDF should repeat the report title, model/version identifier, generation time, confidentiality or workspace context when applicable, and page numbers. Source links must remain readable and clickable where supported.

Safeguards for consequential decisions

Expected value does not authorize trading one person’s safety, rights, or material harm for benefits elsewhere. Apply mandatory constraints before optimization and conduct separate safety, environmental, equity, legal, engineering, financial, and procurement review as required. In construction, validate quantities, costs, geotechnical and schedule assumptions with qualified professionals and current site-specific evidence. In finance, do not treat a decision tree as investment advice, a solvency assessment, or a guarantee of return.

Technical calculation guardrail

Albarena preserves every saved collection record. Schema sanitation limits identifiers to 160 characters; names, titles, owner, and horizon to 500; comparable-state keys to 200; source references and supporting-list entries to 2,000; unit labels to 100; currency codes to 12; and other narrative fields to 8,000. Any longer source is shortened only with an explicit blocking normalization finding that requires comparison with the source. A separate technical preflight pauses derived calculations when the model exceeds 5,000 nodes, 10,000 branches, 5,000 evidence items, 1,000 learning comparisons, 100,000 aggregate supporting-list items (evidence links, assumptions, stakeholders, and constraints), or 256 nodes on a directed path, or when it contains a cycle. This is a computation-safety envelope, not a plan or storage limit. No derived collection is silently shortened, and no partial recommendation is shown. Split or simplify the tree before calculating again.

The interactive Syncfusion diagram has a separate rendering envelope of 250 nodes and 500 branches per view. Exceeding either boundary pauses only the diagram: the complete saved model remains intact and editable through the searchable node inspector, and deterministic calculations can continue while the larger calculation-safety envelope still passes.

Approval and audit trail

Record who reviewed which model version, in what role, when, with what status and reservations. Substantive edits to the tree, values, probabilities, learning costs, or objective invalidate prior endorsement until re-reviewed. Preserve both the model inputs and derived analysis so calculations can be reproduced without relying on private AI transcripts.

AI disclosure

Identify material AI-assisted drafts. AI may summarize the structured model, explain results, or flag gaps, but it may not silently change the authoritative tree, mark evidence as verified, record consultation, approve readiness, or make the final decision. A persuasive narrative never overrides validation warnings or missing evidence.