The Internal Contradictions

Eleven chapters, and the same hole in five of them.

It arrived first as a stale definition: a term defined correctly in 2023, still sitting in the system, still being applied, describing a world that has moved. Then as a formula whose purpose expired, running smoothly, producing consistent answers to a question nobody needs asked anymore. Then as process drift, where the method has changed by increments nobody chose and every instrument for detecting that is built from the process's own recent history, which drifted along with it. Then as captured knowledge going quietly out of date in a searchable place, which is worse than not having it, because a document that can be found is a document that gets trusted. And then as a derivation whose accuracy declines without any signal at all, because a derivation has no way to be surprised.

Five chapters, five subjects, one failure. A thing that was right when it was built, that is still running, that nobody is examining, and whose wrongness produces no symptom distinguishable from its correctness.

I did not plan that. Each chapter went looking for the weaknesses of one ism and four of them found their way to a sixth wall I had not marked on any map. Arriving at the same place from five directions is the strongest evidence available that the place is real rather than an artifact of how I framed any particular question.

So this chapter starts there, because it is the largest thing the creed does not handle, and because every other contradiction in the book is smaller than it.

State it precisely, because the precision is what makes it tractable.

Every rule in this creed converts a decision into a standing answer. Define a term, hold a method, write a formula, record a fact, derive a field: in each case something that was being worked out repeatedly stops being worked out. That conversion is the entire value proposition of the book and I am not retracting it.

A standing answer has a property that a repeated decision does not. It contains, invisibly, the state of the world at the moment it was made. The definition assumed what a warranty period meant in a particular year. The formula assumed which conditions were observable and who could manufacture them. The standard assumed a set of inputs. The derivation assumed a mapping between phone numbers and accounts. None of those assumptions is recorded anywhere, because at the moment of conversion they were not assumptions. They were simply how things were.

The world then changes, at its own pace, without notifying anything. And the standing answer keeps producing output at the same rate, with the same confidence, in the same format, because nothing in it is connected to the thing that changed.

Which produces the property that makes this hard rather than merely annoying: a rule that has gone wrong and a rule that is still right are behaviorally identical from inside the system. Both run. Both produce consistent output. Both pass every conformance test, because conformance tests compare behavior to the rule and the rule is being followed perfectly. The only thing that would distinguish them is a comparison against the world, and the comparison against the world is precisely the work the rule was built to eliminate.

That is not a bug in one of my isms. It is a structural consequence of all of them, and it says something I would rather not say: the creed systematically converts visible ongoing costs into invisible accumulating ones. The visible cost was a person deciding something repeatedly. The invisible cost is a decision that stopped tracking its subject. The first shows up in everybody's accounting. The second shows up nowhere until it shows up as an incident, and the incident will be attributed to whoever was standing nearest.

The instruments each chapter proposed all have the same shape and the same gap, which is itself informative.

The expiry works where the underlying fact carries a date. A license, a certification, a contract term. This is the only instrument that runs unattended, and it covers a small fraction of what goes stale, because most facts have no expiration printed on them.

The override rate works where people can depart from the rule and the departure is recorded. A scheduler who calls a firm the matrix did not list, a clinician who dismisses an alert, an operator who does it a different way: each is the system reporting on itself, and a rising override rate on one rule is that rule telling you it no longer matches the world. Two limits. It requires that overriding be possible and recordable, which many systems deliberately prevent. And it is silent about rules people have stopped consulting, which are the ones furthest gone.

The case that will not fit works earliest and gets discarded most. Somebody forces a request into the nearest category, the forcing resolves the immediate problem, and nothing records that a near miss occurred. The information existed for about four seconds in one person's head.

The manual sample works on anything and costs real time. Pull a handful of instances a month, check them against the world by hand, deliberately and without automation. This is the only instrument with no blind spot in principle, and it is the one every efficiency argument in this book is pointed at eliminating.

Look at what they have in common. Every one of them requires a person to be in contact with the material after the system has removed their reason to be there. The expiry is the only exception and it barely covers anything. The creed's central move creates a problem whose only general solution is to deliberately undo a small part of the central move.

I find that genuinely uncomfortable and I have not found a way around it. What I will say is that the undoing is cheap relative to the conversion. Sampling twenty derived values a month is not a meaningful fraction of what the derivation saved. The reason it does not happen is not cost. It is that a sampling routine looks like waste on every metric an automation project is evaluated on, and nobody defends a line item that exists to discover that the project was wrong.

So the practical form is a budget rather than a principle. Any rule that runs unattended gets a fixed, small, permanent allocation of human attention whose only job is to compare the rule against the world. Not to improve it. To find out whether it is still true. That allocation is part of the cost of the automation and should appear in the case for building it, and if it is not funded, the honest description is that the organization has decided to run the rule until it fails visibly.

The second contradiction is between two isms that point in opposite directions past a certain depth, and I have been getting away with it because I have never applied both at full strength to the same thing.

Define every variable to its deepest identifiable condition pushes down. Don't make the specific unnecessarily vague pushes toward saying what the reader needs. Those sound compatible and they are not, because the first has no terminating condition that comes from a reader and the second has nothing but that.

Chapter 4 asserted that the vagueness rule should win where they conflict, on the grounds that it carries a brake and the depth rule does not. That was an assertion and I owe it a test.

Here is the test. The two rules are answering different questions, and the confusion is that both questions sound like how much detail.

The depth rule governs the structure of a definition: how many separately identifiable conditions it decomposes into, which determines what questions the system can answer and what situations it can distinguish. That is Ashby's floor. A definition with four states cannot tell apart twelve situations, and no amount of clarity in the prose fixes it.

The vagueness rule governs a transmission: what one person tells another in a particular message, which is bounded by what the receiver is going to do next.

Those are different objects. A system can carry a deeply decomposed definition and still produce short messages, because the depth lives in the structure and the message reports only the part that changes the reader's action. That is not a compromise between the two rules. It is what both of them recommend once you notice they are talking about different things.

Where they genuinely collide is at the intake form, and the collision is real. Asking a person for the deeply decomposed version means many fields; telling them only what they need means few. Chapter 12 resolved that by technique: added fields must be cheap, which means chosen rather than typed and derived where possible, so specificity rises while effort does not.

I will defend that as far as it goes and name where it stops. The technique works when the deep structure can be reached by derivation. Where the deep structure can only be reached by asking a person, there is a real trade and no principle resolves it, and the correct move is to accept a shallower definition than the structure would support rather than a form nobody completes. A definition you cannot populate is not a deeper definition. It is a fiction with more fields.

So the collision resolves partially, by separating structure from transmission, and leaves a residue at the intake, and the residue is settled case by case in favor of what people will actually fill in.

The third contradiction is between standardizing and improving, and this one I consider genuinely solved rather than negotiated.

Do the same thing the same way every time and change the way when it can be better are contradictory only if both are supposed to be running continuously. Juran's trilogy separates them in time rather than splitting the difference: control holds the method so the process produces a stable distribution; improvement changes the method as a designed intervention; control then holds the new one. What is forbidden is not change. It is undesigned change, of which there are exactly two kinds, and naming them is what makes the resolution operational.

Tampering is adjustment in response to a fluctuation. Deming's funnel measures the cost: correcting for the last error leaves the drops spread about forty percent wider than doing nothing, which is variance doubled, and the two worse rules diverge without bound. Every organization runs these, and they look like responsiveness.

Drift is nobody deciding anything. It is the failure the previous section is about and it is worse, because tampering at least leaves evidence in the form of somebody's decision.

Two enemies, one permitted change, and the permitted one has to be a decision somebody made and recorded. That works, it is not a fudge, and it is the cleanest resolution in this chapter.

The fourth contradiction is between minimizing what a person does and keeping the person capable, and it does not resolve. It gets priced.

Bainbridge's ironies are correct: the tasks left after automation are the ones that resisted automation, which are the hardest; the operator's remaining job is monitoring, which people are poor at; manual skill degrades unused and is demanded at the worst moment; and the mental model that makes intervention possible was built by doing the thing that was removed.

Chapter 12's distinction between friction and contact narrows this and does not close it. Some input is a tax and should go. Some input is how a person stays in touch with the situation. The three tests offered there help: does filling this require looking at the situation; could the answer be wrong in a way the system cannot check; what happens to the person's model in a year.

But there is no version of this where you get the efficiency and keep the capability for free. The honest statement is that maintaining the capability costs something ongoing, that the something is practice, and that practice means deliberately doing work the system could have done. That is the same shape as the decay budget from the first section, which is not a coincidence: both are permanent allocations of human attention that exist to preserve a property the automation silently consumes.

Which suggests they should be the same allocation, and in practice they can be. A person who samples the system's decisions against the world every month is practicing the judgment and auditing the rule at the same time. That is the only piece of good news in this chapter, and the sample is small by design, and small practice is not the same as competence.

There is a fifth pairing the outline listed and it does not need relitigating here, which is itself worth recording.

We shouldn't pull information, we should push information against attention as the binding constraint was supposed to be one of this chapter's main events. It was settled inside Chapter 11 instead, and settled hard: the rule was rewritten rather than qualified, because Simon's conservation relationship between information and attention is not a competing consideration to be balanced. It is a budget the original sentence did not know existed. Push what changes an action, to the one person whose action it changes, on a channel reserved for that; everything else is made findable.

What survives into this chapter is not the contradiction but its residue, and the residue is a design problem nobody has solved. If the system filters, the filter's mistakes are invisible by construction, because a suppressed message leaves no trace of having been suppressed. If the recipient filters, we are back to a pull system with extra steps, failing the same people for the same reason. The filter has to live in the system, be designed by somebody accountable for what it drops, and have a visible error rate, and the clinical alerting literature is fifty years of very capable people failing at exactly that with better data and higher stakes than anything I work on.

I record it here rather than in the contradictions above because it is not a conflict between two of my rules. It is one rule meeting a fact about people, losing, and being replaced. That happened four times in Part I and it is the healthiest thing in the book.

The fifth contradiction is the one that most threatens the book's own method.

The creed's twelfth rule is identify your bias, then set it aside, and it is the rule this book applies to itself. It says question everything, including your own conclusions, especially the ones you are fond of.

Compiling a formula says stop asking, which is what compiling is. A decision that gets re-examined every time it comes up has not been converted into anything.

Chapter 8 proposed a truce: the questioning applies to the rule and the compiling applies to the execution, so the rule gets audited on a schedule and nobody re-decides it mid-shift. That was a scheduling trick, and Chapter 8 said as much.

It is better than a trick and it is not a full resolution.

What makes it more than a trick is that the two activities have different objects. Executing a rule and evaluating a rule are different acts requiring different information, and running them simultaneously is not rigor but paralysis. A dispatcher who re-derives the scheduling logic on every assignment is not being thoughtful; they are failing to have a process. Separating them in time is the same move Juran's trilogy makes for control and improvement, and it is correct for the same reason.

What keeps it from being a full resolution is that the compiling makes the evaluating harder in a way the separation does not address. A rule you have executed ten thousand times is a rule you can no longer see. This is the human-compilation problem from the repeatability chapter: the compiled version feels identical from the inside because compilation removes the step at which you would have noticed. Scheduling an audit does not restore the perception that compilation removed. It puts a person in front of a rule they have stopped being able to evaluate, on a Tuesday, with a form.

The only partial answer is that the auditor should not be the operator, which is inconvenient, expensive, and the one thing organizations reliably will not do because the operator is the person who understands it. I do not have a better answer. This is the contradiction I would most like somebody to solve, and I am fairly sure it is not solvable by a rule, because any rule for auditing would itself compile.

The sixth is emotion, and it has been open since the first page of this book.

My working position is that emotion is excluded as noise from data-based decisions by default, and admitted as a weighted variable only where the result determines how I will feel. The drive-thru equation exposes the problem: every one of its three terms carries the word perceived, and a perceived cost is a quantity set by feeling. So the terms I excluded emotion from are terms whose values emotion assigns.

That is not a contradiction I can dissolve with a distinction. Both things are true. The arithmetic is real and the weights are not arithmetic.

What I can say, having written eleven chapters since, is where it actually bites, and it is narrower than I feared. It bites on decisions about oneself, where the person setting the weights and the person living with the result are the same and there is no external standard. It bites much less on organizational decisions, not because feeling is absent there but because the weights are contested by multiple people with different feelings, and contested weights have to be stated out loud to be settled, and stated weights are inspectable.

Which produces an odd and I think correct conclusion: a formula is more trustworthy when several people with different interests had to agree on its weights than when one person set them alone, and the personal use of this creed is therefore the harder case rather than the easier one. Everything in this book is easier to apply to an organization than to yourself, and I have been presenting it in the other order.

Three smaller ones, stated and not resolved, because they belong on the record.

The creed does not distinguish automation by whose interest it serves. The mechanism is identical whether the rule was written by the person it runs on or by somebody selling their attention. Chapter 14 offered three questions as a first cut: can the person read the rule, does the released capacity return to them, can they turn it off today at no cost. Those are cheap and they are not a principle. The gap matters because the largest and most effective application of this book's method in the world is aimed at the people the book is addressed to.

The regress in definition has no bottom. Chapter 2 adopts a stopping rule of convenience, borrowed from Deming's three parts: a term is defined when a test, a criterion, and a yes-or-no decision exist. That works and it is a convention rather than a solution, and the honest position is that every definitional system rests on terms somebody agreed to stop questioning.

And the boundary ism produces no evidence of itself. Every other rule builds something you can audit. The rule for leaving things alone builds nothing, so the creed systematically over-reports its own construction and under-reports its own restraint. The manual-decision register proposed in Chapter 15 is the counterweight and it costs about a minute a month, which means it will not be kept, which means the asymmetry is permanent in practice even though it is solvable in principle.

And deep thinking has no definition. Chapter 14 named this against itself: the creed's most important claim uses a term that fails the creed's own first test, because no procedure exists for determining whether an hour was spent on it. That is not a small irony. It means the one ism that says what all the others are for is the one ism that cannot be evaluated, which puts the book's purpose in the position the book spends four chapters warning about. I have no repair for it. What I can say is that the failure is honest rather than evasive: I could manufacture a test, something about hours in uninterrupted blocks on problems without known answers, and the test would be a proxy, and Chapter 8 is a long argument about what happens to proxies that govern anything anyone cares about.

One more, and it is the contradiction between this creed and the person holding it.

Every rule in this book was written by somebody who likes rules. That is not a disqualification and it is a bias with a direction, and the creed's own method says to identify your bias and then set it aside rather than pretend it is absent. So: I find converting a decision into a formula satisfying in a way that is independent of whether the conversion was worth making. The satisfaction arrives when the rule is built, not when the rule turns out to have been right, and those are separated by months or years, which means the reinforcement lands on building rather than on building well.

That has consequences I can see in my own work. I have automated things that did not need automating and experienced it as productivity. I have reached for the boundary ism less often than the evidence in Chapter 15 says I should have. And every instrument this chapter proposes, the sampling and the register and the scheduled evaluation, is a thing I would find tedious, which is a fact about me and predicts something about whether they will be kept.

The reason to write that down is that a creed which does not account for its author's preferences is a creed that will be followed exactly where it is pleasant and abandoned exactly where it is not, and the pattern of that abandonment will look like judgment. Naming the direction of the bias is the only correction available, since I cannot stop having it.

If you are reading this and you are the opposite kind of person, the one who finds rules constraining and prefers to decide fresh each time, then everything in this chapter inverts. Your abandonment will be on the conversion side rather than the maintenance side, and the rule you will overuse is the one holding that where a data-based decision will not produce a genuinely better result, you leave it manual. It will feel like discernment.

What comes out of all of that is not a set of repairs. It is one pattern, showing up six times.

In every contradiction in this chapter, the creed converts something ongoing and visible into something standing and invisible, and the conversion is correct, and the invisible thing then requires a small permanent allocation of exactly the attention the conversion was meant to free.

The decay hole needs sampling. The skill problem needs practice. The compiling-versus-questioning problem needs a scheduled evaluation by somebody who has not compiled it. The boundary ism needs a register. Every one of those is a person, spending time, on purpose, doing something the system made unnecessary.

That is the honest shape of this creed and it is not what the creed says about itself. Freeing the human mind is the promise. The reconciled version is that automation frees a large amount of attention and then reclaims a small amount permanently, as maintenance, forever, and the ratio is excellent and the small amount is not optional.

An organization that takes the freeing and skips the maintenance has not implemented this book. It has implemented the first half of it, which is the half that shows up in the business case, and it will run beautifully for somewhere between one and three years.

Chapter 18 turns the recurring instruments into a single test. Chapter 19 restates the twelve isms with every narrowing visible, and it reconciles against what the chapters actually argued rather than against what I expected them to argue, which in at least four cases was different.