Hallucination From First Principles: Building Reliable Systems Around Unreliable Models

Ernan Hughes

Aug 2026 · Ernan Hughes

About this ebook

Large language models can produce answers that are fluent, confident, and wrong. The difficult question is not simply why hallucination happens, but how to measure it, detect it, evaluate our detectors, and stop unsupported claims from becoming accepted state.

Hallucination From First Principles develops a practical engineering framework for doing exactly that.

Rather than treating hallucination as a single mysterious defect, the book breaks it into measurable failure modes and builds a reliability system from the ground up. It separates truth from evidence, support, attribution, provenance, verification, and policy acceptance—distinctions that become essential once language models are used in real systems.

The book examines what can be measured directly, what requires external evidence, and where apparently strong detection methods fail. It explores embedding geometry and Hallucination Energy, claim decomposition, retrieval, natural-language inference, LLM judges, consistency testing, adversarial evaluation, abstention, calibration, verification, and the control of persistent memory.

Along the way, you will learn why:

• fluent generation can continue after evidential support disappears

• a good benchmark score does not guarantee a reliable detector

• stronger reasoning models can still hallucinate

• consistency is not the same as truth

• retrieval is only useful when evidence actually supports the claim

• LLM judges require evaluation rather than trust

• thresholds should reflect the cost of different errors

• verification and policy become more important as model outputs gain consequences

The book does not claim that hallucination has been solved, or that any single detector is sufficient. Instead, it develops a methodology for combining imperfect signals, discovering their boundaries, and deciding what model output should be accepted, rejected, verified, or allowed to persist.

The objective is not perfect generation.

It is controlled acceptance: a system that expects its model to fail, knows how to measure the failure, and has explicit rules for deciding what survives.

Written for engineers, researchers, and technically curious readers, Hallucination From First Principles moves from the mechanics of generation to the architecture of reliable AI systems.

It begins with hallucination.

It ends with reliability.