Front Matter: Hallucination, Mechanically

Hallucination, Mechanically treats AI hallucination as an engineering problem: unsupported claims moving through systems without enough evidence, authority, or mitigation.

Key Takeaways

• The book defines hallucination mechanically: a system emits unsupported, contradicted, stale, or action-inventing claims.
• The recurring motif is evidence binding: claim to source span, source span to authority, authority to mitigation.
• CLAIM turns hallucination reduction into a pipeline: decompose, link, check authority, test inference, then decide whether to answer.
• The useful output is not a slogan; it is an operating checklist for retrieval, citation, summarization, agents, abstention, evals, and production.

Read this with the opening litigation failure, the CLAIM Framework, and the evals field guide.

Book promise

Hallucination is not one bug with one fix. It is a family of distinct failures that happen to share a symptom: a system makes a claim that no available evidence supports, and says it fluently. This book takes that family apart. It separates fluent guessing from retrieval failure, retrieval failure from unsupported synthesis, unsupported synthesis from stale knowledge, stale knowledge from bad evaluation, and bad evaluation from product design that never gave the system a way to say I don't know. For each failure mode it shows the mechanism, how to detect it, what actually reduces it, what only appears to, and what stays risky.

This is a practical, mechanistic field manual for the people who ship AI answers: AI engineers, product engineers, MLOps engineers, technical founders, support and search teams, and the engineering leaders who own the on-call rotation when the system tells a customer something untrue. It is not a philosophy of truth, not a RAG tutorial, not a benchmark leaderboard, and not a prompt list. It is a systems manual for reducing unsupported claims, written with the assumption that you have already shipped something that hallucinated, watched it happen, and want to know which failure it was.

The recurring motif

**Fluency is not evidence. **

A language model is trained to produce coherent continuations of text. Coherence is the objective; truth is, at best, correlated with it through the training data. When the model writes a sentence that sounds careful, measured, and certain, that tone is a property of the language, not of the facts. The model can be exactly as fluent about a real court case as about an imaginary one, about a number that appears in the source as about a number it invented to fill the slot. Every chapter in this book is, in some form, an elaboration of the gap between sounding right and being supported.

The enemy

The belief this book exists to correct:

"LLMs hallucinate", said as if hallucination were a single behavior with a single cure, so that lowering the temperature, adding retrieval, asking for citations, or buying a bigger model should fix it.

Sometimes those moves help. Often they fail, because the actual cause was something else: the right document was never retrieved; it was retrieved and ignored; two sources contradicted each other; the reasoning ran past the evidence; the model's parametric knowledge was a year stale; the task was ambiguous; a tool returned an error that the model narrated as success; or the evaluation rewarded plausible prose and so trained the team to ship plausible prose. You cannot fix a failure you have not named.

Primary research references

These anchor the book. Individual chapters use their own chapter-specific sources; this is the shared spine.

• Survey of Hallucination in Natural Language Generation (Ji et al.)
• A Survey on Hallucination in Large Language Models (Huang et al.)
• TruthfulQA: Measuring How Models Mimic Human Falsehoods
• Language Models (Mostly) Know What They Know
• SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection
• Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks (Lewis et al.)
• RAGAS: Automated Evaluation of Retrieval Augmented Generation
• Lost in the Middle: How Language Models Use Long Contexts
• FActScore: Fine-grained Atomic Evaluation of Factual Precision
• OWASP Top 10 for LLM Applications

The CLAIM Framework

One framework recurs through the book. Whenever the system is about to assert something, ask five questions:

• **C: Claim. ** What factual statement, exactly, is the system making? Decompose fluent prose into atomic, checkable claims.
• **L: Link. ** Which specific source span, tool result, or state record supports this claim? Not the document, the span.
• **A: Authority. ** Is that source allowed for this user, current as of now, and authoritative for this question? A real citation to a superseded source still produces a wrong answer.
• **I: Inference. ** Did the model stay inside the evidence, or did it infer beyond it? Unsupported synthesis is the failure that retrieval cannot fix.
• **M: Mitigation. ** Given the answers above, should the system answer, revise, ask a clarifying question, abstain, or escalate to a human?

CLAIM is used as a lens, not a template. It will not appear as a forced subsection in every chapter. It is the question set a mature anti-hallucination system can answer for any sentence it emits.

Table of contents

Movement I: Naming the Failure

1. The Confident Wrong Answer
2. A Working Taxonomy of Hallucination
3. The CLAIM Framework

Movement II: Why Models Produce Unsupported Claims

4. Fluency Is Not Evidence
5. What Models Know About What They Know

Movement III: Retrieval Failure Is Not Generation Failure

6. When Retrieval Fails Before Generation Begins
7. A Citation Is Not Proof

Movement IV: Summaries and Transformations

8. The Compression Press

Movement V: Tools, Agents, and Hallucinated Actions

9. Hallucinated Actions

Movement VI: Detection and Verification

10. Claim Extraction and Source-Span Verification
11. Self-Consistency and the Limits of the Judge

Movement VII: Mitigation Patterns

12. Interventions and Their Limits
13. Teaching a System to Say "I Don't Know"

Movement VIII: Evaluation and Monitoring in Production

14. Measuring Unsupported Claims
15. Operating Against Hallucination in Production

Movement IX: Use-Case Playbooks

16. Playbooks by Domain

Back matter

• Glossary
• Implementation Checklist
• Research and Source Register

The practical entry point for new readers is Introduction: The Holding That Never Was, which opens with a concrete failure and sets up the book's diagnostic approach.