Book

Description

Large language models can produce fluent, persuasive, and internally coherent answers even when the evidence needed to support them is incomplete or entirely absent. This book argues that such failures are not merely problems of data, retrieval, or verification. They point to a deeper limitation in how these systems are functionally organized. Drawing on neurological research into confabulation and the left-hemisphere interpreter, the Neural Arbitration Framework introduces a conceptual architecture built around two complementary modes: a generative mode that constructs responses and an epistemic mode that identifies uncertainty, contradiction, and the boundaries of reliable knowledge. An arbitration protocol governs the relationship between them, allowing the epistemic process to interrupt generation before uncertainty hardens into a confident but unsupported answer. The book positions this proposal alongside retrieval-augmented generation, post-generation verification, uncertainty estimation, and multi-agent approaches. It also sets out possible implementation pathways, falsifiable predictions, and a cognitive self-mapping mechanism through which a system could record recurring limitations and turn them into a structured agenda for improvement. Rather than presenting a finished technical solution, this book asks a more fundamental question: can artificial intelligence learn not only how to answer, but also when an answer should not be given?