Corpus-Guided Discovery of Computational Structural Invariants: Representational Deficits and Auditable Knowledge Admission

Knowledge-based systems need explicit criteria for admitting candidate knowledge and governing changes to evolving representations. This study presents an auditable framework for corpus-guided discovery of candidate computational structural invariants, evaluated through a longitudinal lexical-engineering study using a frozen historical Portuguese corpus. Its central component is a knowledge-admission layer that distinguishes retrieval, relation formation, structural correspondence, abstraction, candidate-invariant formation, representational audit, implementation, and validation as separate epistemic states. The framework prevents similarity from being treated as an established relation and candidate isomorphism from being treated as validated knowledge. It also recognizes the absence of a match, unresolved cases, and decisions against promotion as legitimate outcomes. Controlled lexical expansion increased recognized occurrence coverage and lexical depth per verse, while adversarial collision cases prevented promotion based solely on similarity. A read-only schema audit identified missing explicit representations for typed lexical-entry relations and lexical families, indicating the need for a candidate schema extension without altering the core representation. A subsequent gate distinguished lexical identity from grammatical realization, while the Portuguese grammar module remained unimplemented. Corpus-wide structural retrieval produced candidate families but did not establish analogues, counterexamples, or isomorphisms. The contribution is an auditable knowledge-admission discipline that connects corpus-derived hypotheses to controlled representational change or the explicit identification of representational deficits. The reported findings are supported by frozen artifacts, formal gates, cryptographic hashes, metric contracts, and published experimental outcomes, without relying on undisclosed production internals.