A claim that cannot lose is marketing wearing a serious face.
The last chapter asked leaders to answer for the parts of a system they cannot hand off. I have spent twenty-one chapters asking companies and people to expose their work to tests they do not control, and it would be cheap to end without turning the same demand on myself. If I want you to test your company, you need to know what would make this book retreat.
I originally wanted to place three registered bets here, complete with thresholds, frozen populations, coders, judges, and a public archive. The protocols were not ready to carry the authority the prose gave them. A private draft is not a public record. An intended judge is not an appointed independent judge. A baseline I calculated is not a verified baseline. So those bets are deferred. Removing them is not a retreat from falsifiability. It is the standard of this book applied to its author: no artifact, no claim.
What remains are three failure conditions already inside the argument. They are not registered predictions and they should not be mistaken for results. They are the places where evidence can press hard enough to change the next edition.
The first is the company thesis. This book argues that AI grades the operating conditions it meets, and that durable transformation requires more than buying tools and inserting them into unchanged work. The argument weakens if companies repeatedly produce durable, enterprise-level results by doing exactly that. Not a demo. Not license activation. Not hours claimed in a survey. A sustained operating or financial result, attributable with reasonable confidence to a deployment placed into an operating model whose direction, data, process, human experience, authority, and evidence practice did not materially change. If that pattern becomes common, the central diagnosis in Movement I is too strong. The company did not need to become inheritable first. The tool was enough.
The hard word is unchanged. It would be easy for me to protect the thesis by calling every successful deployment evidence that the company must have changed somewhere. That would make the argument impossible to lose. So the burden runs the other way. A future test would have to define operational change before examining the outcome, preserve the evidence used to make the call, and allow an independent reviewer to disagree. Until that protocol exists, this is a disconfirmer to watch, not a result I can claim to have tested.
The second is the framework. The five questions are offered as a candidate inspection grammar for the knowledge-work scope this book defines. The proposal weakens if a relevant organization repeatedly presents no meaningful referent for one or more questions. Its operating logic weakens if prospective uses identify a weakest evidence-supported condition, then readable outcomes repeatedly show that the condition did not constrain the result. And the framework should retire if a rival instrument guides decisions more reliably with less burden, less ambiguity, or less risk of misuse. A framework does not earn permanence because its author gave the parts memorable names.
Notice the standard. An explanation written after the outcome does not count. The state has to be recorded before the result is known. Rival explanations stay attached. No reading remains a legitimate outcome. The person settling the result cannot depend on my reputation or approval. Those requirements make a future public test slower. They also make it worth reading.
The third is the Blank Collar itself. The name is useful only if it preserves the two-move practice. Repeatable work becomes inheritable. Then the person grows into accountable work the inherited system cannot settle. Directing a system of people and agents without verifying its results and owning its consequences does not meet that definition. If the term travels only as another name for an AI user, a flexible employee, a generalist, an agent manager, or a person who simply produces more, then the category has failed even if the words become popular. Distribution without meaning is not a win. It is a slogan escaping its argument. The person-side compact can fail too. If organizations provide legitimate authority, learning, transition support, recognition, and credible next responsibility, and people who transfer repeatable work still systematically lose the ability to contribute rather than grow into new responsibility, then the compact is incomplete. The handoff may improve continuity while failing the person. Evidence may show that even both moves are not enough.
None of these conditions gives me a number to announce today, and that is the point. The honest sequence is artifact first, public protocol second, claim third. If a verified public test is later opened, it should show its definitions, dates, hashes, conflicts, missing evidence, and independent decisions before it asks anyone to trust the outcome. Until then, this chapter does not borrow authority from an archive that does not exist.
You do not need to wait for any of that to act on Monday. One workflow can still be inspected. One repeatable responsibility can still be made inheritable. One person can still receive a real next responsibility with the authority to carry it. The book can be wrong at the level of its largest claims and still owe you an instrument honest enough to discover where. That is the standard I am willing to be held to.