How to Read This Book

This is a field manual for someone who wants to find out what actually happened, using a machine to help, and who intends to do it honestly enough that the finding survives contact with a court, a commander, or a journalist.

It is not a book about how clever detectives are. There is a version of this subject that flatters the reader — the one where a brilliant mind looks at a scuffed shoe and names the murderer. That version is the reason so much real investigation goes wrong, and the first chapter takes it apart before anything else can be built.

The spine

Everything here hangs from one sentence:

Investigation fails where a human narrows too early. The machine's real gift is not intelligence — it is that it removes the need to narrow at all. And that gift is void the moment its output is treated as an answer instead of a lead.

Both halves matter. A book that only sold the machine would be marketing. A book that only feared it would be useless to anyone with a hundred thousand case files and eleven officers. The evidence in these pages supports both halves, and it is presented in a form that lets you check.

The four grades

Every substantive claim in this book carries one of four tags. They are not decoration. They are the book's central argument in miniature.

The fourth grade is the unusual one. Most books quietly skip what they could not find. This one prints it, because in this subject the gaps are frequently more informative than the data. Thailand publishes no clearance rate for online fraud. That is not a hole in the research; it is a fact about the system, and this book treats it as one.

These four are not a stylistic choice. Three of them are, almost exactly, the four-valued logic Nuel Belnap proposed in 1977 for a machine being fed conflicting reports by multiple sources: told true, told false, told both, told neither. [ESTABLISHED] says told-true; [CONTESTED] says told-both; [ABSENT] says told-neither. The fourth, [CLAIMED], has no counterpart in Belnap's scheme, because his machine trusts its sources and investigation cannot — it records not what was said but who had an interest in saying it.

Why that matters, and why it is more than a curiosity, is worked out in full in an appendix at the end. The short version: a system that expresses all four of those states as a single confidence number cannot distinguish we have evidence against this from we know nothing about this — and one of the documented disasters in these pages is precisely that confusion, running at scale, on real people.

Here is the asymmetry that made the grading necessary. Almost every story you will encounter in which artificial intelligence catches a criminal is [CLAIMED] — sourced to the company that sold the software or the force that bought it. Almost every story in which it harms an innocent person is [ESTABLISHED] — sourced to a court record, a settlement, or an inspector general. That asymmetry does not prove the technology fails. It proves that failures generate public documents and successes generate press releases. Read every enthusiastic number in this field with that in mind, including the strongest ones I found.

What each chapter owes you

Every chapter closes with two things.

The Move — one named, runnable action. Not a principle. Something you can do on a Tuesday.

What would change my mind — the observation that would break the chapter's claim. A chapter that cannot name its own falsifier is an opinion wearing a citation.

What this book deliberately does not do

It does not teach you inference. It does not explain Bayes' theorem, catalogue cognitive biases, or argue at length about why "deduction" is the wrong word for what detectives do — though the first chapter says so in passing, because the correction cannot be skipped entirely.

That work is done in the companion volume, The Science of Deduction / ​ศาสตร์​แห่ง​การอนุมาน, which sits in the same library. That book is about the reasoner. This one is about the job: what closes cases, what the laboratory can bear, how separate reports get joined into a network, where deployed systems have hurt people, and how to do open-source work without committing a crime.

Where the two touch, this book compresses and points. It does not repeat.

What it will cost you to use

Two things, and they are the price of everything in here.

First, you must be willing to write down what would prove you wrong before you look for what would prove you right. This is the single behaviour the research supports most strongly and the one practitioners skip most reliably.

Second, you must keep the machine's output in the category of lead rather than finding, in writing, every time, including the times when it is obviously correct. Every documented disaster in this book is a case where someone stopped doing this once.