What the Test Suite Could Not See: An Error Taxonomy from Building a Neuro-Symbolic Food-Label Compliance System for Arabic Labels

Abstract

AbstractFood-label compliance verification is a high-stakes regulatory task, and the neuro-symbolic pattern — a probabilistic perception layer feeding a deterministic judgment layer — is an established response to it. This paper does not propose that architecture. It reports what happened when a single domain expert without a programming background built one, over fifteen months, for Arabic-language food labels checked against Jordanian national standards and Codex Alimentarius provisions.The contribution is an error corpus rather than a system. Ninety-five malfunctions were recorded in a development log at the time they occurred, each anchored to a commit, a named product, or a quoted system output. They are classified here under a seven-class scheme frozen before coding began, along a second axis recording what actually surfaced each one.The central observation lies in the second axis. The project's automated regression suite — twenty-two frozen input-output cases, green throughout — surfaced one of the ninety-five incidents, and none of the sixteen perception failures. Perception failures were surfaced overwhelmingly by one route: a domain expert holding the physical package. Sixteen further incidents were not system failures at all but false assertions about the project's own state, made by handoff documents and by the language model assisting the build, and were surfaced almost exclusively by live inspection of the executing code.We report these as counts of documented incidents, not as rates. The corpus is a record of what was noticed by one observer who was also the system's builder, and no detection rate can be derived from it.Keywords: neuro-symbolic AI, regulatory compliance, food labelling, Arabic NLP, error taxonomy, experience report, LLM-assisted development