Catelingo: Constraint-Based Semantic Validity Verification for Language Models

Large language models frequently produce outputs that are syntactically fluent yet semantically invalid.
While recent verification approaches focus on reasoning chains or factual knowledge, many semantic failures occur without explicit reasoning or missing facts.
Such errors include temporal inconsistencies, numerical impossibilities, and semantic type clashes, which cannot be reliably detected by reasoning-based or knowledge-based methods.

This paper introduces Catelingo, a constraint-based semantic validity verifier that detects these failures by checking the satisfiability of explicit semantic constraints.
Rather than evaluating truth, likelihood, or factual correctness, Catelingo defines semantic validity as constraint satisfiability induced by an input sentence.
The proposed approach requires neither reasoning chains nor knowledge retrieval, and does not rely on model retraining.

We implement a toy version of Catelingo using a small, sense-level lexicon and explicit constraint propagation over syntactic dependency structures.
Experiments demonstrate that Catelingo correctly detects semantic no-go cases involving temporal ordering violations, numerical range violations, and semantic type incompatibilities.
We further show that metaphorical expressions can be selectively permitted through explicit degeneration rules, and that domain adaptation can be achieved by switching constraint profiles rather than retraining models.

These results suggest that constraint-based semantic verification provides a lightweight and scalable complement to existing reasoning- and knowledge-based verification methods, addressing a class of semantic failures that they do not cover.
An open-source reference implementation and deterministic test cases are provided for reproducibility.
The source code for the verification, Catelingo, is open-sourced at https://github.com/ShinobuMiya/Catelingo

This work is intended as a design-oriented technical report and is not tied to a specific benchmark or leaderboard.