Neuro-symbolic Formal Verification of Culinary instructions via ontology-grounded constraint satisfaction

Abstract

This study proposes the «Formal-of-Thought» (FoT) approach—an architecture that shifts artificial intelligence from passive generation to active formal verification using the FoodOn ontology. FoT uses large language models (LLMs) as a specification compiler, «mapping» DenseNet-121 features to FoodOn proofs using the «Retrieval-Augmented Generation» method. Safety constraints are formulated as logical predicates, and compliance with them is verified using an SMT solver.