Overture — Fifteen Minutes

We have built machines that produce evidence. We have not yet built the discipline that allows the rest of us to trust them.

A courtroom in rural Wisconsin, March 2026. The clock above the bench reads 3:47 in the afternoon, which means the parent on the witness stand has been sworn for two hours and seventeen minutes, and the State's expert — a mid-career social worker with a doctorate in family-systems counseling and an impressive list of trainings — has just told the court that the parent has missed seventeen scheduled visits with a six-year-old child since the previous February. The expert is reading from a printout of the agency's case-management software. The printout is one column of dates and one column of red Xs.

The parent's attorney is forty-one years old and was appointed two weeks ago. She has read the file. She has not read the parent's phone.

The judge looks at her over a pair of half-moon glasses and says: if you wish to challenge any of these dates, do so now. We are not coming back to this.

Her client leans forward and whispers, hoarsely, that ten of those seventeen were rescheduled by the agency. They told me, the parent says. By text. The day before. Sometimes the morning of.

The attorney has the parent's phone. She has fifteen minutes. The phone is a six-year-old iPhone with twelve years of accumulated text messages, including conversations with at least thirty-seven different case managers, nine of whom no longer work for the agency. The conversations interleave with messages from family, employers, the mortgage company, three therapists who came and went, and several thousand transactions of unsolicited marketing. The contacts list never canonicalized the case-manager phone numbers. They appear under at least four different identifiers per manager — work cell, personal cell, an extension that was reassigned, and a name with no number.

She searches reschedule. One match. She searches moved. None. She searches can't make it. Six matches, none from the agency. She flips through pages of conversations by hand. Her client watches her, and the clock, and the judge.

Time runs out.

The expert's testimony stands. The parent's testimony to the contrary stands without corroboration. The proceeding moves forward on the record as it now exists, which is a record in which seventeen visits were missed and the parent was not, in any sense the law recognizes, present in the child's life.

This book is not about that proceeding.

This book is about a quieter fact. The text messages were on the phone the whole time. The agency's records, parsed correctly and searched correctly, would have shown the rescheduling. So would the parent's email, which contained automated calendar updates from the agency's scheduling system. So would the parent's calendar, which had been updated each time the new appointment was offered. The truth was distributed across six private systems, none of which spoke to any of the others, and fifteen minutes was not enough time, by any human or any tool the parent had access to, to assemble it.

This is a problem of engineering.

It is also a problem of law. And of journalism. And of the regulators who will, in the next decade, decide what evidence may be machine-assembled and what burden of disclosure attaches to it. And of every person whose digital life has accumulated, over the past decade, into a corpus larger than they can read, which is, by now, almost everyone.

Solving it requires three things to hold together at once.

The corpus must be comprehensive — every record the case might turn on has to be in the system, demonstrably and exhaustively. It must be accurate — the system's interpretation of what each record says must be correct in a sense an attorney or auditor can defend on the record. And it must be verifiable — every output the system produces must be checkable, by an outside party, without trusting either the system or its operator.

Every existing tool achieves one or two of these. None of them, until recently, achieved all three.

This book is about the discipline that does. The discipline has a name — the Canon — and a technical specification, and a reference implementation, and a set of cryptographic primitives borrowed from twenty years of work on signed-software supply chains and content authenticity. None of that vocabulary will be necessary in the first few pages. By the end you will know all of it.

What you will also know, at the end, is what to do the next time you walk into a room with fifteen minutes and a phone full of evidence and a judge looking at you over half-moon glasses.

You will not be the parent's attorney that afternoon. You will be the person who built the system that, three years earlier, would have made that afternoon a different one.

This book has five readers, and you are at least one of them.

You may be a software engineer — a builder of systems — and the material in this book is your craft. You will read every chapter and do every lab. The book will give you a working architecture for evidence systems, the cryptographic primitives that make them defensible, and the discipline (gap disclosure, applied-and-declined challenges, falsifiability without issuer cooperation) that is, today, almost completely absent from the software industry's normal practice.

You may be a lawyer, a judge, a regulator, an auditor, on whose desk these systems are starting to land as evidence and exhibits and expert reports. You don't have to learn the cryptography. You will read the framings, the worked examples, the legal appendices, and the sidebars that explain what to ask of the engineer who built the system in front of you. By the end you will know the questions a court should ask of machine-generated evidence, and how to evaluate the answers.

You may be a journalist or a researcher working with private archives — FOIA productions, leaks, oral-history corpora, whistleblower documents — that conventional tools cannot index reliably. The vocabulary in this book is yours; the techniques are yours; you may finish it having decided to build a small Meridian of your own. The chapter on the Investigator's reading path tells you which sections to focus on.

You may be a policymaker or a standards staffer, watching as the courts, the agencies, and the legislatures begin to grapple with what evidence means when an AI is in the loop. This book is the engineering brief that, until now, has been missing from your committee record. The chapters on FRE 901, FRE 707, the Daubert framework, and the EU AI Act overlay are written for you. So is the chapter on the Admissibility Auditor — a deterministic check that organizes the questions a court must ask, into a form a court can act on.

You may, finally, be the person whose evidence the system handles. A pro-se litigant. A subject of an investigation. The beneficiary of an estate. A person whose private records are being decided over by people who have not read them. The book does not promise you an outcome. It promises you the vocabulary to ask the right questions of the people who will: was the system that produced this exhibit verifiable without the issuer's help? what challenges did it apply, and which did it decline, and why? if it makes a claim about my conduct, where in the artifact does the falsifier live?

A book that addresses all five of you is unusual because the discipline it teaches has no existing constituency — engineers do not yet know they are responsible for it, lawyers do not yet know how to ask for it, journalists do not yet know it is available, and the affected do not yet know the language to demand it.

The constituency is forming. Forming it requires that all five of you read the same volume.

Three observations carried me, over five years, to writing this book.

The first is that an extraordinary fraction of contested factual disputes — civil, criminal, regulatory, journalistic — now turns on records that exist only in private digital systems. Email, instant messages, voice notes, GPS breadcrumbs, payment ledgers. We do not have the procedural infrastructure for handling these records that the paper-records era took two centuries to build. The bonded document examiner. The notary. The clerk of court. The sealed envelope.

The second is that AI-assisted retrieval is both indispensable and unverifiable. Indispensable because no human will ever again read the seven hundred thousand items in a typical decade-old digital archive. Unverifiable because no human, certainly no judge or jury, can meaningfully introspect what a 137-million-parameter neural network does when it ranks a document third instead of seventeenth.

The third is that the legal system is, right now, in the early stage of working out what to do about this. Mata v. Avianca in 2023 was the opening of the line of cases that established attorneys are responsible for the AI tools they use. Park v. Kim in 2024 made it Second Circuit precedent that Federal Rule of Civil Procedure 11 reaches AI hallucinations without any new rule. Johnson v. Dunn in July 2025 escalated to public reprimand and disqualification: a court declared that "modest fines have proven insufficient." Federal Rule of Evidence 707 — the proposed rule that would specifically govern machine-generated evidence — finished its public-comment period in February 2026. The earliest it can become operative law is December 1, 2027.

We are inside the window. The discipline this book teaches is the discipline a court in 2027 will, on an early estimate, expect.

Some of what follows is dense. Cryptography is, at certain depths, mathematical, and the math is not optional. But every dense section is twinned, in this book, with an In the Wild sidebar — a real case, a news story, a courtroom moment — that makes the technique tactile. And every chapter has a Why It Matters box for the reader who wants to know what is at stake before they decide whether to read the next twenty pages.

You can read the book end to end. You can also read it as a builder, or as a lawyer, or as a journalist, or as a policymaker, or as the person the evidence is about. The reading-paths matrix at the front of the book shows you which chapters to read in which order for which role.

You will learn, by the end, the same five things, in some sequence:

What a cryptographic hash is, and why it converts a document into a detectable record of its own integrity.
What a digital signature is, and how it converts an integrity record into a binding commitment by a specific issuer.
What canonicalization is, and why two systems can agree on every field of a document and still disagree on its bytes — and what that disagreement makes possible.
What an attestation is — the four-block (Witness, Findings, Refutation, Seal) artifact that is the central object of the book — and how the declined-challenge inventory it carries is the closest thing the field has to a falsifiability discipline.
What it means to verify without the issuer's cooperation, and why that property — and not any number of certifications, standards, or endorsements — is what an evidence record in 2027 will be measured by.

You do not need to commit to the whole book to begin. Read on. The parent's attorney comes back, in three years, in a form you can recognize. She is a different person; the parent is a different parent; the case is a different case. The fifteen minutes is the same.

This time she is ready.