A hybrid light-AI-based framework for accessibility validation

Automated accessibility evaluation systems increasingly leverage large language models (LLMs) to assess conformance with established standards like the Web Content Accessibility Guidelines (WCAG). However, a fundamental methodological tension exists between the probabilistic, generalization-based nature of LLM inference and the deterministic, precisely scoped requirements of normative accessibility standards. A trustworthy, transparent, and accountable accessibility validation framework must be grounded in three principles: normative anchoring of evaluation logic to official guideline definitions, deterministic and traceable execution of assessment checks, and the systematic production of verifiable evidential artifacts. To operationalize these principles, we propose a computationally lightweight first, hybrid methodological framework for accessibility validation. The framework enforces a logical structure indicating the use of AI components, including small language models (SLMs), small multimodal models (SMMs), vision models, graph representations, and web agents, when rule-based techniques are insufficient to reliably evaluate criteria requiring human-like judgment and contextual understanding. We demonstrate the utility of this methodology, highlighting how it ensures algorithmic transparency and auditability, through concrete case studies.