Prefailure Transformer Representation Transitions: A Falsifiable Hypothesis for Mathematical Change Before Commitment and Observable Factual Failure

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Pre-Failure Transformer Representation Transitions investigates whether an AI model begins to go wrong internally before the failure becomes visible in its final answer. We developed a mathematical framework for measuring changes in the model’s internal representations and identified a measurable transition point where those representations begin to drift from their stable behavior. The work establishes that visible failure may be preceded by an observable internal warning signal, creating the foundation for detecting and potentially intervening in AI failure before the model produces the wrong output.