Use AI as a proposal layer

The assistant receives a bounded, normalized context—not an unrestricted raw snapshot. It may critique the frame, suggest missing dimensions or values, identify possible pair tensions, draft typed constraints, surface evidence gaps, or critique a candidate.

AI responses use a structured proposal envelope. Applying a proposal is explicit, non-deleting, and identity-safe: the parser rejects attempts to overwrite protected identifiers, computed evaluations, confirmation fields, verification fields, or human promotion. New AI pair assessments remain unconfirmed. New AI constraints remain advisory. AI evidence remains unverified. AI cannot reject configurations or select a strategy.

Review what changed before applying. Ask the assistant to distinguish facts, user-supplied information, inference, assumptions, and items needing verification. If context was truncated, use the workbench’s targeted action for the relevant dimension, pair, configuration, or candidate rather than assuming the assistant saw the whole model.