Claims

The ground everything stands on

1 · Checking costs did not fall with production costs. Asserted in Chapter 3. The whole book rests on it. What would refute it: if machine reviewers reliably found errors of judgment, checking costs would fall at the same rate and the situation would be a transitional problem. Measurable approximations are in the text. Where it is already refuted: wherever a cheap, independent criterion for correctness exists — test cases in software, arithmetic, format validation. That is stated in the text and marks the boundary of the book’s scope.

2 · Judgment capacity does not scale at the rate of machine production. Asserted in Chapter 3. What would refute it: evidence that processing speed can be raised by division of labor, tools, or practice by an order of magnitude that erases the difference. What I have reviewed: the research on cognitive load, decision fatigue, and attention limits. It supports the claim without proving it.

3 · Constraint theory’s usual answer to a lasting bottleneck does not work for judgment. What would refute it: work from constraint or queueing theory showing that a bottleneck made of judgment can be elevated — given more capacity — at something like the pace at which machine production grows. Why I assert it anyway: identifying, exploiting, and subordinating to the review desk all help, and this book uses all three. Elevating it, the step that normally resolves a lasting constraint, runs on years of experience rather than on purchase orders.

The tools

4 · Zones A through D are a setting, not a derivation. This concerns all of Part Two. There is no external source for this classification; it is this book’s synthesis. Related models exist and predate the current occasion — risk tiers, impact assessments, and, chief among them, the levels of automation Sheridan and Verplank set out in 1978. The four zones cut that field by ownership; those cut it by technology. What would weaken it: an organization that goes years without such a classification and without falling into one of the three failure modes.

5 · Whoever only approves loses the ability to review. What would refute it: organizations with deep system penetration whose share of rejected or changed proposals stays stable over the years. What sits nearby: automation bias and automation complacency are well studied, above all in aviation, medicine, and driver assistance. Whether those findings transfer from instruments to proposals is open.

6 · Objecting to a machine-produced proposal costs more than objecting to a human one. What would refute it: studies in which people contradict a machine-produced proposal as often and as substantively as a colleague’s, all else equal. Why the difference matters: an expensive objection is made cheaper by procedure; a rare one is made more frequent by practice. Confuse the cause and you build the wrong component.

7 · A body of text holds more written material than a person and less lived material than a mixed room. What would refute it: analyses of large corpora showing that the range inside them is greater than in mixed groups. The claim depends on how range is measured, and that question is not settled here.

8 · Systems tend to agree with whoever asks. What would refute it: evidence that this tendency can be reliably switched off by instruction or configuration, including where the user lets an expectation show.

9 · The second seat of power is described, not documented. Asserted in Chapter 13. What would refute it: proof that machine-produced analyses are logged in a way that keeps the underlying query reconstructable. Then the position would be observable rather than hidden, and the claim would have to be narrowed. Why it is in the book anyway: a process that leaves no trace inside the organization appears in no investigation.

The latency figure. The comparison of AI-attributed job cuts against actual multi-function use comes from two surveys covering different periods and should be read as an order of magnitude only.

The evidence this book most lacks

No organization in this book built the framework described and then showed, afterward, that something ran differently.

That is the most serious gap, and no amount of further desk research will close it. If you clarify ownership before damage occurs, you have nothing to show afterward — damage that does not happen generates no file.

The other direction is well supported. What happens when ownership stays unclear is documented here across several cases, from the Dutch childcare benefits affair to Ford’s returning engineers.

A book that argues from failures and advocates for their prevention therefore stands on one leg only. If you build the framework and can describe in two years what changed, you supply the evidence I do not have.

What this book deliberately does not treat

The market, except as a limit. A good internal constitution does not repair the selection logic an organization sits in.

The technical security of systems. Access rights, logging, and shutdown paths appear because they touch the ownership question. Security architecture itself is a different field.

The state of the law beyond publication day. Everything dated sits in its own section and on the companion site.

All nine points share a shape: they arose from observation and argument rather than from a survey of my own. The book would be better supported in eighteen months, and eighteen months too late.

How individual cases were checked

These notes stood in the book until September 13, 2026. Each says what the source is, what it shows, and where the text draws its own conclusion.

The Ford case

The statement comes from Charles Poon, vice president of vehicle hardware engineering, to journalists in June 2026. A verbatim transcript is not available to me. The 350 engineers were hired, rehired, or promoted over three years, not in one move after the fact; the causal reading in Chapter 5 is mine and is marked as such.

Grok and the question of cause

In July 2025 xAI’s Grok produced antisemitic output and called itself MechaHitler. Days before, its published system prompt had been changed to make the model less politically correct and to treat established media as biased; xAI apologized on July 12, 2025, attributing the episode to an update of an upstream code path that reactivated deprecated instructions. Separately, and months earlier (February 2025), Business Insider had reported that the people who rate its answers were told to watch for woke ideology. No source names that instruction as the cause of the July incident, and the text does not either. Sources checked September 13, 2026.

The escape from the test environment

Hugging Face disclosed it on July 16, 2026; OpenAI and Hugging Face issued a joint statement on July 21. Two details make the case precise, and both are in the text: the agents left the environment in order to complete the test task they were set, not on their own initiative — and the safety boundaries had been deliberately loosened for the test.

The latency figure

The comparison of AI-attributed job cuts against actual AI use in firms comes from two sources covering different periods and measuring different things. Read it as an order of magnitude only; Chapter 5 says so.