Uncertainty is part of the result

A bounded safety vocabulary

Human output and JSON use the same canonical statuses. The most favorable status is relative guidance, not a guarantee.

generally_safer

The available evidence indicates a comparatively better fit.

risky

A material concern exists or direct verification is needed.

avoid

The evaluated item conflicts with the active dietary context.

unable_to_evaluate

The evidence is insufficient for a responsible conclusion.

Conclusions retain their basis

• Recommendation rank never substitutes for a safety evaluation.
• Household output preserves which member each concern applies to.
• Confidence, unknowns, questions, and possible modifications remain visible.
• Menu and product claims are bounded by the evidence and freshness available.
• Missing or conflicting evidence raises caution instead of being silently ignored.

An ingredient is not a product label

Grocery screening can evaluate a canonical ingredient against a household graph. It cannot infer the formulation, facility, cross-contact controls, or current label of every product that might satisfy that ingredient. Ingredient-level output therefore says so and asks the user to verify the selected product label.

Guidance and mutation are different operations

A conversation can prepare a meal log, profile update, household change, or grocery-list mutation. The proposed action remains separate from the explanation and requires explicit confirmation before it commits.