The Revelation Threshold

On the Friday, four days before the meeting, Maria showed it to somebody.

Not to anyone who would be in the room. To the man from logistics — the one who said the thing about pallet labelling at the first Improvement Time session four years ago, which she is still using, and whose name she had to think about for a second before she said it. They ended up at the same table in the canteen. He asked how the deviation reports were going, in the way people ask about work they do not do, and she heard herself say that she had a way of doing them now. He looked interested. She offered to show him.

They went back to her desk after lunch, and before she could open anything she had to say what it was.

She found she had no way of saying it. Nine months, and the thing had never had a name. In her own head it is the report thing, and on her laptop it is a folder called dev with a date after it, and neither of those survives being spoken out loud to a colleague standing behind your chair. So she said: it is a little tool I put together for the deviation reports. That is true, and it is the first sentence anybody in that building has ever heard about it, and she chose it in about a second and a half.

Then she showed him the wrong thing.

Not by accident and not because she meant to hide anything. She opened the output rather than the tool, which is the summary it produces on a Friday morning, and she talked him through that, and twice she said the word rough before he had said anything at all. When he asked how it decided what to put at the top, she said it looks for repeats. Which is not what it does. It clusters by failure mode across suppliers, and the reason that is the whole point is that the standard report is organised by supplier, and explaining why that ordering is wrong would have taken ten minutes and would have amounted to saying that the way this plant has read its deviation data for as long as anyone can remember is the wrong way round. That is a criticism of a system, and a system belongs to people, and two of them sit on her floor.

So she showed him a faster version of the report he already knew.

He was impressed by it, which was worse than if he had not been, because what impressed him was the wrong thing — the speed. He said it must save her hours. She said it did. He asked whether it was on the network and she said no, it is just on here, and he nodded in a way that closed the subject, and she registered that he had understood something she had not said.

Then he said the dangerous thing, and he said it kindly.

He said she should show it to the man with excellence in his job title. Not as a warning and not as a hint. As the obvious next step, offered by somebody being helpful, because that is what you do with a thing that works — you take it to the person whose job is things that work. He is right by every ordinary standard, and Maria has spent eight chapters arriving at the reason he is wrong, and she said something about maybe, after the review, and he let it go.

On the way out he asked whether anyone else knew.

She said no. He said she should tell them. And then the conversation moved to something about a trailer that had been sitting in the yard for three days, and they walked back through the corridor with the poster in it, and nothing further was said about any of this.

What changed that afternoon was not the tool. The tool is exactly what it was at half past eleven. What changed is that there is now a person in the building who could say one sentence about it in a room she is not in — not out of malice, not even out of carelessness, just conversationally, in a meeting where somebody asks whether anyone is doing anything about deviation reporting. Chapter 9 was about what a bad surfacing does to the person for the rest of their career. This is the term before that one. Before the sanction can be applied there has to be something to apply it to, and until Friday lunchtime there wasn’t.

That evening she did something she has not done in nine months, which is write the decision down. Two columns on the back of a printed agenda.

She could not fill in the left one, because Chapter 6 already told you what is in that box and it is not knowable yet. The right one she could fill in, and the first thing she wrote at the top of it was not they say no.

It was: they take it.

There is a formal decision rule for exactly this, published last year in an innovation journal, with terms and a threshold and a small amount of algebra. Maria arrived at it in a canteen without writing anything down, which is roughly the relationship this whole book has to the literature.

The rule, already written down

The cleanest statement of the surface-or-submerge decision is not mine, and it is barely a year old. In 2025 Julian Waters-Lynch, Darcy Allen, Jason Potts and Chris Berg published a paper on what they call shadow user innovation — covert use of consumer AI tools inside firms, positioned deliberately at the junction of the user-innovation literature, the shadow-IT literature, and the bootlegging work this book opened with. It is the most on-point external source I have found for the whole subject, and it does something none of the older literatures do. It writes the employee’s choice as a rule.

Their model has two triggers and one threshold. Covert use starts where there is acute pain in a work process and where the fix is concealable. It ends — that is, the person discloses — when the expected gain from disclosing exceeds the expected sanction risk. Below that line they stay covert. They call the line a revelation threshold.

Read as a description of a person it is almost insultingly obvious, and that is a recommendation rather than a complaint. Good formalisations of familiar behaviour usually look obvious. What the formalisation buys you is that it separates the decision into terms, and terms can be examined one at a time, and the moment you examine them one at a time against the previous nine chapters something becomes visible that is not visible in the prose version.

The rule is correct. Every input to it is wrong in the same direction.

Four terms, all bent the same way

Take them in order, and notice as we go that none of what follows requires anybody in Maria’s plant to be unreasonable.

The gain, at the moment of asking. The expected benefit of disclosure has to be estimated by the person disclosing, on the day they disclose. Chapter 6 spent itself on why that estimate is systematically low: the value of a good increment is produced by running it, in the place it runs, over a period, and the largest part of it is usually something nobody predicted. Maria in month zero could not have priced the seal defect because in month zero the tool did not catch seal defects. So the gain term is not merely uncertain. It is biased downward by construction, and biased hardest for the increments most worth having, which is the paradox from Chapter 6 restated as an input to somebody’s arithmetic.

The gain, conditional on a yes. Here is where the equation quietly assumes something false. It reads as though a favourable response converts the expected gain into the actual gain — as though the yes delivers the thing. Chapter 8 established that it does not. The yes arrives from a person who is good at his job, wants the plant to do well, and is structurally unable to receive a small working thing except by giving it a sponsor, a scope and a name. The smallest version of Maria’s tool that the organisation is capable of adopting is larger than the tool. So the gain has to be discounted not by the probability of rejection but by the probability that acceptance is of a different object.

The sanction, which does not discount for being right. Chapter 2 gave the sharpest version of this and it is worth repeating, because it is the finding that most people refuse to believe until they have watched it. Allen and colleagues found that employees who used unsanctioned tools were marked down whether the tool succeeded or failed. The penalty attached to the circumvention. A rational person facing that evidence cannot reduce their sanction estimate by building something excellent, because excellence is not the variable being punished, which means the one lever the engineer actually controls is disconnected from the term it ought to move.

The sanction, which does not expire. And this is Chapter 9 arriving as a coefficient. A bad surfacing is not a one-off cost paid on a Tuesday. It is an update to a belief the person will carry into every subsequent decision of this kind, for years, without ever re-testing it, because the self-editing runs before the conversation starts. Maria at twenty-six paid this once. What she has been carrying since is not a memory of a stalled proposal but a permanently raised sanction term, and it is raised for every idea she will ever have, not for the one she lost.

Four terms. Two of them understate the gain, two of them overstate or extend the risk, and all four errors push the same way — down, toward concealment.

Which gives the fleet its real explanation, and it is not the one most executives arrive with. The submarine is not crewed by people who are frightened, or disengaged, or insufficiently included in the vision. It is crewed by people doing correct arithmetic on inputs the organisation has bent.

That is a considerably worse problem than a morale problem, because you cannot fix it by being nicer to anybody. But it is at least a problem with a shape. If the terms are bent by the machinery, then the machinery can in principle be unbent, and Chapter 12 is an attempt at that.

There is one more thing wrong with the equation, though, and it is not a bent input. It is a missing term, and it is the reason this chapter exists rather than being a paragraph in the last one.

The curtain

I have been holding this since Chapter 1, and it is the strongest piece of evidence in the book.

I should say plainly what I mean by that, because this book has been careful about the difference between an analogy and a result. Almost everything in Parts II and III arrives here from somewhere else. The capital research is about plants and equipment. The attention research is about subsidiaries. The selection research is largely about chief executives of listed companies, and I flagged the size of that leap where I made it. Tyre and von Hippel and the NUMMI study are the two bricks that sit on an actual assembly line, and they are case studies.

What follows is a controlled result, obtained on a production line, in a working factory, by a researcher who went in looking for the opposite of what he found. There is nothing else like it in this book and I am not going to pretend the rest of my evidence is of the same kind.

In 2012 Ethan Bernstein published a paper in Administrative Science Quarterly based on work inside the world’s second-largest mobile phone factory. He went in to study transparency, which was then and is now one of the few management ideas nobody argues with — that the more of the work managers can see, the better the work will be. Visibility improves quality. Visibility catches error. Visibility spreads good practice from the line that has it to the line that does not.

What he found on those lines was that the operators had worked out better ways of doing the job than the official ones. Faster, and in some cases more reliable. Improvements of exactly the kind the previous nine chapters have been about, invented in place, by the people with their hands on the work, without any of it having been asked for.

They hid them.

When a manager came down the line, the improvised methods went away and the operators returned to the procedure in the manual, and they went back to the better version once the coast was clear. This was not occasional and it was not a few troublemakers. It was how the floor worked. And the suggestion boxes standing on every line, which existed precisely so that this would not have to happen, stayed empty — which Chapter 9 has already explained, and which at the time must have looked like a workforce with nothing to say.

Then Bernstein hung a curtain.

Not a metaphor. Fabric, around parts of the production floor, creating zones the operators could work inside without being observed from outside — with the output still measured, and the line still accountable for what it produced. He reduced what managers could see, on purpose, in a factory, and left everything else in place.

Performance improved, and it stayed improved.

I want to be exact about the direction of that, because it is the part people reverse when they retell it. The gain did not come from the workers being happier, or trusted, or motivated by a gesture. It came from the improvements being allowed to exist. The better methods already existed before the curtain; what the curtain removed was the reason to put them away, and the line got the benefit of work its own people had already done and had been hiding for as long as anyone had been watching.

Bernstein named the finding himself, and his name for it is the one this book will use. He called it the transparency paradox.

His prescription is narrower than the story usually told about it, and the narrowness matters enough that I will not smooth it over. He is not arguing for factories nobody looks at. He argues for bounded privacy, with specific boundaries — privacy around teams rather than individuals, a separation between feedback and evaluation, a separation between the right to decide and the right to improve, and protected periods for experimentation. Observation is not the enemy in his account. Observation wired directly to judgement is. That distinction is doing quiet work here and it will do loud work in Chapter 12.

The principle, and where you have already met it

Now go back and look at what happened in the canteen.

Maria was not afraid of the man from logistics. He has no authority over her, he had already been helpful to her once, and nothing about that afternoon carried any risk that would show up in the equation above. And within ninety seconds of him standing behind her chair she had opened the output instead of the tool, described the clustering as it looks for repeats, apologised for the state of it twice, and presented the whole thing as a faster version of the report the plant already has. She showed him the by-the-book version. She had a curtain and she took it down herself, and what came out from behind it was the manual.

That is Bernstein’s operators, in an office, with a laptop, nine months in.

And it is the third time this has happened in this book. The first was in Chapter 6, at the form. Expected benefit (quantified) did not measure Maria’s tool badly. It measured it accurately, in month zero, when it clustered by supplier and saved her an afternoon — and month zero is the least the thing will ever be worth. The looking did not misread the value. The looking chose the moment, and the moment set the value.

The second was Chapter 7, at programme scale. Nobody killed the protected time by judging it and finding it wanting. Somebody sensibly asked how many people were using it, and somebody sensibly asked what had come of it, and each request was reasonable, and together they rebuilt the gate inside the sanctuary. The audit did not discover that the programme had stopped being a nursery. The audit is what stopped it being one.

Three altitudes — an idea, a programme, a person — and one mechanism running at all three.

Observation does not reveal the state of a hidden improvement. It sets it. And it sets it downward.

That is the second organising claim in this book, alongside the three-currency convergence from Chapter 5, and I will call it the observation principle, on the understanding that the name is a container for Bernstein’s finding rather than a rival to it. He established the mechanism at the scale where it can be measured. The claim that the same mechanism is operating at the other two scales is mine, it is an argument rather than a result, and the caveats section deals with how far it runs past him.

But notice what it buys, because it is more than tidiness. Everything in Part II mis-valued the increment downward, and never once upward, and that asymmetry has been sitting there unexplained since Chapter 4. Here is the explanation. Looking early collapses a thing to whatever is legible on the day of looking, and legibility correlates with size, with precedent, and with distance from the actual work — which is Chapter 6’s closing line about a system that quietly optimises for forecastability and calls it rigour. The direction of the error is not an accident of any particular gate. It follows from the act of measuring an embryo.

Four results, and none of them is the same study

I have just built an organising claim on one factory in one country, and the caveats section will say so again in harder terms. Before it does, it is worth laying out what else is standing near it, because the situation is better than one study and worse than a settled literature, and the difference between those two matters.

Four results, none of which was designed to test any of the others.

A factory. Bernstein’s lines, where privacy raised output and the effect held for months. A field experiment with a control, in the setting where the work is most visible and most standardised.

A laboratory. Amabile’s 1990 crossing of observation and evaluation, twenty-two years earlier, where the expectation of judgement — not the watching — reduced the creativity of what people produced. Different method, different decade, different participants, no organisation anywhere in it.

A workgroup. Brown and Baer’s territorial marking, from Chapter 4, where the act of one party staking a claim measurably suppressed the creativity of everyone else in the room. Nobody rejected anything. The environment changed and the output changed with it.

A decision meeting. Mueller’s bias against creativity, also from Chapter 4, where the evaluator’s own uncertainty degraded the reception of exactly the ideas most worth having, and did it below the level at which the evaluator could feel it happening.

What they share is not a hypothesis. It is a direction. In all four the conditions of being seen or judged altered what was produced or offered, rather than revealing what was already there, and in all four the alteration ran the same way — toward the safe, the conventional, the defensible. Four methods, four scales, one direction.

What that does not do is prove the observation principle. These are not four tests of one claim, and I am not going to present them as if they were; three of the four were run for other reasons and would not have been designed this way. Convergence of this kind is weaker than replication and stronger than a single site. What it buys is a specific and limited thing: if Bernstein’s factory turned out tomorrow to be a fluke, the direction would still have three independent results behind it, and the argument would be wounded rather than dead.

What that does to the equation

The revelation threshold treats the improvement as a constant. There is a thing, it has a value, and the decision is whether to reveal it — with the same object sitting on both branches, unaffected by which branch is taken.

The observation principle says that is exactly what it is not. Disclosure is not a report about the tool. It is an event in the tool’s life, and it is an event that changes what the tool is: what gets shown, how it gets described, what it will be permitted to become afterwards. Maria’s tool on the concealed branch clusters by failure mode and catches things five weeks early. Maria’s tool on the disclosed branch, described in twenty minutes to somebody standing behind her chair, is a faster report — and if the room adopts anything on Tuesday, that is what it adopts, because that is what was in the room.

Two branches, two different objects. So the comparison the equation asks for cannot be made in the form it asks for it.

Which does not make the threshold useless. It makes it a rule about timing rather than a rule about worth, and that is the more useful thing to have. The question stops being is this good enough to show and becomes has this become what it is going to be, and the second question, unlike the first, has an answer that changes with the calendar.

So the question has a date on it

Chapter 6 explained why delayed assessment works, and it is worth pulling the practical consequence out of it, because nine Surface Checks have now said not yet and I have never said what the yet is waiting for.

Criscuolo, Salter and Ter Wal found that hidden projects came out better, and the mechanism was time — the embryo gets to become what it is going to be before anyone judges it. Nothing in that finding says hidden forever. It says hidden until, and the paper is not in a position to tell you when the until arrives, because it measured outcomes rather than moments.

The practitioner literature is more direct about it and less rigorous, which is the usual trade. Miller and Wedell-Wedellsborg’s Innovation as Usual, and I will flag that as a trade book rather than a study, argues for stealth in the early stages on the grounds that a premature no is close to irreversible — once a proposal has been declined, everything subsequently learned about it gets read through the decline, and the people who declined it have a position to defend. Their advice is to stay under the radar until the thing can survive contact.

Put that against the observation principle and the shape of an answer appears, though not a comfortable one. If observation sets the value, and it sets it at whatever is legible on the day, then the only thing under the discloser’s control is which day. Not whether to be seen. When. And the honest test for the day is whether the thing has accumulated something legible that is also true — a rate you can defend from your own records, a second person who has used it and would say so, an event it has already survived. Maria has the first of those, has had it since about month four, and has neither of the others.

The Surface Checks have been saying not yet for nine chapters. What they have been waiting for is a version of the tool that can be seen without becoming worse, and nothing in her plant is building one, and she cannot build it alone, because two of the three things on that list require somebody else.

The decision in its real shape

Part IV opened by promising to make the dilemma rigorous, and the rigour has turned out to cut in an awkward direction. The decision rule is real and it is not the book’s. Waters-Lynch and colleagues wrote it down properly: disclose when the expected gain beats the sanction risk, otherwise stay under. Everything this book has assembled goes into that rule as inputs, and every one of the inputs is bent the same way. The gain is estimated on the day the value has not yet been produced. The gain on a yes is discounted by the fact that a yes converts the object into a larger one. The sanction does not fall when the work is good, because the penalty attaches to the circumvention. And the sanction, once paid, is not paid once.

Then the principle takes the last piece of solid ground away, which is the assumption that there is one object being decided about. There is not. There is a thing that exists in the dark and a different, poorer, more legible thing that exists in the light, and the act of moving between them is what makes the second one out of the first.

So Maria’s arithmetic is sound and her conclusion is correct and both of those facts are a catastrophe for the plant, because the plant needs the tool and there is no move available to her that delivers it. That is not a failure of nerve on her part and it is not a failure of goodwill on anybody else’s. It is a system in which the only way to see the thing is to spoil it.

Which is where the executive enters, holding the obvious solution, and it is the worst idea in this book.

If people conceal, look harder. If the concealment is invisible, get better instruments — and there has never been an instrument for this like the one every large company has just finished buying. That instinct is not stupid, it follows directly from everything above, and it is precisely inverted. The observation principle is now on the table as a principle rather than as a story about a curtain, and Chapter 11 is what it predicts.

Where the evidence stops

The load in this chapter is carried unevenly, and the unevenness is worth mapping before somebody maps it for me.

Bernstein is one factory. It is a Chinese contract-manufacturing plant, on assembly lines, with a defined standard method that operators were deviating from — which is the cleanest possible setting for this finding and not obviously the same setting as a quality engineer with a laptop and no standard method to deviate from. The step from an operator hiding a hand movement to Maria hiding a clustering rule is a step. I think it is a short one, because the thing being concealed in both cases is a local improvement to a prescribed process and the trigger in both cases is being watched by someone who could act on what they see. But it is one site, one industry, one country, and I am leaning on it harder than any other single source in the book. The convergence section above is what stops that being fatal, and it is worth being clear that it does not stop it being a weakness.

His prescription is narrower than my use of it. Bernstein argues for bounded privacy with four specific boundaries, not for reduced visibility as a general good, and executives have entirely legitimate reasons to see things — capital, risk, compliance, coordination. Where this chapter says observation sets the outcome, his finding says something more precise: observation wired to evaluation and control sets the outcome. I have tried to keep that distinction visible, and Chapter 12 depends on it completely, because a resolution built on the loose version would be a resolution nobody could operate.

The revelation threshold is a model, not a measurement. The Waters-Lynch paper is a theory piece. It proposes a decision rule and derives testable propositions from it; it does not report a test. Nobody has measured a revelation threshold in a working organisation, established what actually sits in the terms, or checked whether people behave as the rule says they do. I have used it as scaffolding for exactly that reason — it organises the argument, and I have been careful not to write as though it settles anything.

Miller and Wedell-Wedellsborg is a trade book, said once already and repeated because it is the kind of source that gets promoted to a study by a helpful reader. The stealth-innovation advice is practitioner reasoning drawn from consulting work. The confirmation-bias mechanism behind it is well established; the specific advice about timing is not tested.

And the three-scale claim is mine. Bernstein established the person scale. Criscuolo and colleagues supply the programme scale indirectly, in the finding that the same bootlegging behaviour associated with less innovation during periods of process formalisation — same behaviour, observed surroundings, inverted result — which is suggestive and is not the same as having been tested as a mechanism. The idea scale, Chapter 6’s form, has no controlled result behind it at all. It has an argument and a good one. Nobody has assembled these as one principle because nobody has had reason to, and if the assembly is wrong, it is wrong in the way arguments are wrong rather than in the way findings are wrong, which is to say you can dispute it without new data.

Here is the study that would settle it. Take a large organisation with a population of tools people have built for themselves — every big company now has hundreds, and most of them do not know it. For each one, record two dates: when it was built, and when somebody with authority first saw it. Then look at all of them a year later and ask which still exist, which are used by more than one person, and what they are doing. If the tools that were shown in month one are in the same state a year on as the tools that were shown in month nine, then the observation principle is wrong outside a production line, Bernstein’s result does not travel, and this chapter is a nice story about a curtain.

My expectation is that early disclosure predicts one of two endings — the tool stops, or the tool becomes a project and then stops — and that late disclosure predicts survival.

The obvious objection to that study is the one I would raise myself, and it is fatal to the cheap version of it. People disclose early because the thing is already good, so any raw comparison measures the tools rather than the timing. To get anything out of it the timing has to be assigned rather than chosen, which means a real experiment inside a real company, which means somebody senior agreeing to keep half of it deliberately unobserved for a year. That is a hard thing to ask for. It is also exactly what Bernstein got somebody to agree to in 2012, which is the other reason that study is worth more than everything else I have.

Next: Chapter 11 — the instinct every executive has at this point, and what the evidence says it costs.