A book that has spent thirteen chapters warning against confident inference from incomplete evidence owes the reader a straight account of its own footing before it starts issuing instructions.
So here is the account, and then here is the instruction, and the second does not depend on the first being settled.
What this is
A theory. Not a law and not a finding — an explanation that fits a large body of published research, a much smaller body of direct measurement, and one pattern that has held for sixty years, and which makes predictions that could turn out to be wrong.
Four tiers. They are set out in one place here because thirteen separate sets of caveats are easy to lose track of, and the total is more useful than the instalments.
Established, and not by me. Bootleg effort raises innovative performance, and the mechanism is delayed assessment rather than avoided assessment. Monitoring produces no reliable performance gain, across ninety-four studies and more than twenty-three thousand people. Expecting to be evaluated degrades creative work, which Amabile established in a laboratory in 1990. Observation degrades performance on a production line, which is Bernstein’s controlled result with a curtain in it. Attention at headquarters is allocated by structural weight and there is only so much of it. Voice extinguishes through futility faster and far more durably than through fear. And capital processes are built to guard irreversible decisions, then route reversible ones through the same machinery because it is the only machinery there is. None of that is in dispute and none of it is mine.
Mine, and well supported. That those mechanisms converge — that capital, attention and career incentives all bend the same way against the same class of small improvement. No single literature says this, because no single literature is looking at all three at once. And the observation principle: that one effect operates at the scale of the idea, of the programme and of the person. That is a synthesis across three bodies of work rather than a finding inside any of them, and I have been careful to say so each time it appears.
Inferred from published models, and not measured. That concealment relocates rather than stops when the cost of visibility rises. An enforcement record from another industry shows that behaviour under prohibition; nobody has measured it under monitoring, and those are different stimuli. Also the two-conditions frame — hide to survive birth, surface to survive adulthood — which is a reading of what the birth-condition research implies rather than something anybody has tested.
Conjecture, with a test attached. That AI is the first wave to move the cost of building and the cost of watching in the same purchase, and that the branch we end up on is therefore decided by wiring. Chapter 13 sets out the study that would settle it and names the result that would falsify it. Nobody has run it.
That is the whole of the footing. Somewhere between a strong hypothesis and a working theory, and closer to the first than I would like it to be.
Why you should act on it anyway
The obvious objection is that a theory in that condition does not license instructions.
It licenses some and not others, and the line falls exactly where this book has been putting it since Chapter 5. An action that is cheap and reversible does not need proof behind it. It needs to be plausibly useful, cheaply abandoned, and harmless if the theory turns out to be wrong. That is a two-way door, and the correct treatment of a two-way door is to walk through it and find out, because the finding out is cheaper than the deliberating. An action that is expensive and hard to undo needs a great deal more, because you only get to be wrong about it once.
Everything I am about to ask you to do is a two-way door. None of it costs real money, none of it needs approval above a single manager, and all of it can be stopped inside a week with nothing lost but the week.
The thing I am asking you not to do is the one-way door. Buying an instrument and pointing it at your own people is expensive, visible and effectively permanent in its effects — the people who went quiet do not come back when the policy is withdrawn, which is Chapter 9’s mechanism and among the better-established things in this book.
I want to name that asymmetry before somebody else does. I am asking you to take small reversible actions on incomplete evidence while asking you not to take one large irreversible action on the same incomplete evidence. That is not inconsistency. It is what a two-way door is for, and if everything else in this book turns out to be wrong, that second half is still the right way to treat a decision of that shape.
First, the thing not to do
I have just told you that something valuable is hidden inside your organisation and roughly where it lives. The obvious use of that information is the worst move available to you.
Do not go and find it.
Not with a discovery report, not with a licensed platform that logs individual prompts, not with a well-meant survey asking people what they have built. What you find will be the version that survived being found, you will lose the rest of it permanently, and the compliance number on your report will improve the entire time this is happening. That is Chapter 11 and it is the sentence I would keep if I could only keep one.
Then, in this order
Five things. The order is part of the argument, and each step is cheaper than the one that follows it.
One. Take the instrument off the table. If a discovery or monitoring capability is on a roadmap in your area, the first move is to stop it, and the second is to find out what question it was bought to answer — because there is usually a real one underneath, and it usually has an aggregate answer. How much of this is happening, in which functions, on what data. Every one of those survives being answered at function level with no names anywhere near it.
Two. Wire it before you buy anything. If something is coming regardless of what you want, then it arrives configured, and the configuration is the whole game. Individual-level output goes to nobody who writes, contributes to or signs off a performance review. Reporting lands at function or site level. No names in the appendix. Three sentences in a configuration document, sent before lunch, and it is the entire difference between a developmental instrument and an evaluative one.
Three. Open one route. This is the same act at three altitudes, and you do the version your seat allows. If you built something, get one other person using it — the only survivability condition available without anybody’s permission. If people report to you, write down what somebody may change without asking and circulate it, which most organisations have never done, so the answer defaults to nothing. If you sign things, grant one Protected Window: six weeks, a cap you can authorise alone, exempt from assessment, one conversation at the end.
Four. Answer people. Ten working days, every time, including no. The response is the deliverable and the decision is secondary, because a no with a reason attached preserves somebody’s willingness to raise the next thing and silence does not. This is the cheapest item on the list and the one with the longest reach, and it is the one most likely to be agreed to and then not done.
Five. Then look — and only by invitation. After the other four, and not before. Ask one person who is not on any list and does not report to you to show you something they made for themselves. Then wait, and do not fill the silence. What comes back will be smaller than the thing, because Chapter 10 says it must be, and it will still be more than any instrument would have found you.
If I am wrong
Suppose the optimistic branch is the right one and the pattern is dissolving on its own, because building got cheap enough that nobody needs to hide anything any more.
Count what following this list will have cost you. Some people got answers within ten days. A few six-week windows ran, and some of them produced something and some produced nothing, which is what a window is for. Somebody wrote down what their team may change without asking. And an instrument that would have been unnecessary under that branch anyway did not get bought.
There is no version of that branch in which any of it was a mistake, which is the property I was selecting for when I chose these five and not others.
And if the pattern holds, you will have built the only thing that makes any difference to it, which is a route — one path from a working thing somebody made on a Thursday to a thing the organisation owns, that does not run through a business case. Most large organisations do not have one. Nobody has ever been asked to build one, because nobody has been able to see what it was for.
You can see what it is for now. That is the whole of what this book was written to do.