Auditing Token Padding and Token Accounting in Large Language Models: A Constraint-Adherence Metric and a Multi-Provider Empirical Study

The AI Token Auditing Protocol (ATAP) introduces a framework for evaluating the consistency and transparency of token accounting in large language model (LLM) services. As organizations increasingly rely on token-based pricing for AI applications, independent methods for assessing reported token usage become important for building user trust and supporting accountability. This paper presents the motivation behind ATAP, describes the proposed auditing methodology, and discusses its potential applications for AI transparency, benchmarking, and consumer protection. The framework is intended as a foundation for future research on standardized approaches to evaluating token accounting across AI systems. Code, data, and adapters available at https://github.com/tflux2011/atap