Bounded Agency in a Deployed Label-Reading System for Allergen Avoidance and Dietary Self-Management

We demonstrate ANONSCAN (name anonymized for review), a deployed mobile system that reads the ingredient declaration of a packaged product from a barcode or a photograph and returns a grounded, plain-language assessment at the point of purchase in about two seconds. For food allergy, strict avoidance is the primary management strategy, making the label decision effectively a clinical decision made without a clinician present. Our organizing principle runs against the prevailing direction of agentic design: rather than granting a model latitude and auditing its output afterwards, we confine generative models to three roles bounded by the shape of their output space, two of them deployed and the third specified, under a single invariant: a model may transcribe a number printed on a package, but no model may compute one. Every derived quantity is required to come from deterministic code; we audit the shipped build against that requirement and report the one violation identified by the audit. Measuring the deployed system on a released 200-product corpus shows why this constraint matters: a majority of allergen verdicts rest on a single ingredient token, extraction fails at whole-panel granularity so redundant declaration confers no protection, with a verdict false-negative rate of $40.8%$ compared with $30.8%$ predicted under independent per-token failure, and no inexpensive signal separates a failed read from a clean one, indicating that the verdict surface requires three states rather than two. Taken together, these measurements locate the binding constraint in database coverage rather than in model capability. At the session, the invariant is checkable live: any attendee can compare a number on screen against the package in their hand, then rescale the serving size and verify the arithmetic that follows.