"Almost everything in life can be reduced to a quantifiable formula."

When it is cold, the user wears a coat.

That is the whole thing, and it is so small it barely looks like a claim. There is an input, a threshold, and an output. Nobody writes it down. Everybody runs it.

Now put two people in the same house with the same thermostat.

One of them is cold at sixty-eight and the other is fine until sixty-two, and they have been having the same argument for eleven years without either of them noticing that they agree completely about the rule. If it is cold, wear a coat. Turn the heat up. Neither has ever disputed a word of it. What they disagree about is a number that appears nowhere in the sentence, because the sentence contains a word instead of a number, and the word means something different inside each of their heads.

Nothing about that argument is a communication problem. Saying it again more slowly will not help, and saying it louder has already been tried.

Move the same structure to a building where forty people work and the stakes change without the shape changing at all. A request arrives. Somebody decides what state it is in, which determines who touches it next, which determines how fast it moves and whether anyone is watching it. The rule that makes that decision exists. It runs dozens of times a week. It has never been written down, it lives in three or four people's judgment, and each of them learned it from a different person during a different year.

Then one of them takes another job.

For a few weeks nothing happens. There is no outage and no backlog. The other two absorb the work and the queue keeps moving. When the noticing finally comes it arrives sideways: a request that should have moved sat for nine days, and somebody asks why. The answer is that it was in a state nobody had put it in. The person who would have put it there was gone, and the person who inherited the queue read the same facts and reached a different conclusion, which was also defensible.

Nobody did anything wrong. The instinct in that room is to look for the mistake and there isn't one. Two people applied judgment to identical facts and their judgment differed by an amount neither of them could see, because neither had ever been asked to say out loud what threshold they were applying.

Then the replacement gets hired and trained by whichever of the two is free that week. Whatever version of the rule that person carries becomes the version the building now runs. The other version does not get overruled. It stops existing. A year later there are three people again, all confident, and the rule has drifted somewhere nobody chose to move it.

A slogan does not run. A rule runs, and a rule has named parts.

The variables are the conditions the rule reads, and the action it produces. Temperature and coat. Request attributes and status. Every one of them has to be a thing somebody can actually observe, which turns out to be the hardest requirement in the sentence.

The quantity minimized is the number of times a person has to decide something that has already been decided. Not the number of decisions. The number of re-decisions, which is a different and much larger figure. Every time the same question is answered from scratch, the answer can come out differently, and the variance is the cost.

The failure signal is two people, same inputs, different outputs. That is the observable symptom, and it is the only one you get. The rule does not look like it is failing. Results scatter for no reason anybody can name, and everybody involved is certain they did it right.

The domain of validity is narrow. The conditions have to be observable, they have to be enumerable, and the mapping from condition to action has to hold still for at least as long as the rule runs. Break any of the three and the formula does not merely underperform. It produces confident wrong answers, which is worse than no rule at all, because a confident wrong answer stops anyone from looking.

And then there is almost. Almost everything in life can be reduced to a quantifiable formula. I put that word in the sentence years ago and I have spent a long time not examining it. It is carrying every exception I have not enumerated. The chapter on where a data-based decision will not produce a genuinely better result, so you leave it manual, is the one that has to make almost do honest work.

None of this starts with organizations. It starts with what a nervous system does.

William James, writing about habit in 1890, would not let the word stay moral. He treated habit as a property of materials before he treated it as a property of people, and his term for it was plasticity, which he defined as a structure weak enough to yield to an influence but strong enough not to yield all at once. A structure that yields to nothing cannot learn. One that yields completely cannot hold what it learned. The interesting behavior happens in the narrow band between, and James is describing physics, not character.

What plasticity buys, in his account, is that habit simplifies the movements needed to get a result, makes them more accurate, and reduces fatigue. Three returns from one mechanism. That is the argument for a written procedure, made about forty years before anybody used the phrase standard work, and made about human tissue rather than about factories.

Ann Graybiel found the machinery a century later. Her work on the basal ganglia describes chunking: the brain packages a sequence of actions into a single unit so that a long routine runs as one item instead of as a string of separate decisions. Her lab later watched it happen. Neurons fire at the start of a learned routine, go quiet while it runs, and fire again when it ends, marking the boundaries of something the brain has decided is worth keeping as a package.

A routine learned well enough has edges. Something marks where it begins, something marks where it stops, and in the middle nothing is deciding anything.

Herbert Simon took the same observation into organizations and supplied the vocabulary for claiming that almost everything in life can be reduced to a quantifiable formula. He distinguished programmed decisions, the repetitive ones an organization can handle with a rule, from nonprogrammed decisions, the novel ones that require judgment. His point was not that programmed decisions are inferior. His point was that attention is the scarce resource, and that every decision you can convert from the second category to the first buys back attention for the decisions that cannot be converted. He also insisted that human rationality is bounded, meaning we do not optimize, we satisfice. We take the first answer good enough to proceed with. A written rule is a way of making the good-enough answer consistent instead of leaving it to whoever is tired that afternoon.

Peter Gollwitzer supplied the form. His implementation intentions restate if A = B, then C = D in a psychologist's notation: decide in advance that if situation X arises, then I will do Y. What makes it work, in his account, is that specifying the situation hands control of the behavior to the situation. The cue does the triggering instead of the person. Automaticity gets manufactured deliberately instead of accumulating by accident.

Wendy Wood's work explains why that handoff sticks. Behavior turns out to be cued by context far more than it is steered by intention, and a habit is triggered by the situation in which it was learned even after the goal that motivated it has faded. Which means a rule placed into a system is not filing. It is the installation of a cue that will fire whether or not the person remembers agreeing to it.

In a world of abundant information, Simon argued, the scarce resource is not data and not labor. It is attention, which cannot be manufactured. An organization has a fixed supply of it per day and spends the supply on whatever it is asked to decide. Every decision converted into a rule returns some of that supply. Every decision left to judgment spends it. That is the whole ledger, and it means a written rule is not a convenience. It is the only way to buy back the one input nobody can produce more of.

And when the formula has more than one thing to weigh, Ralph Keeney and Howard Raiffa wrote down how that works. When the attributes come in different units, you make them comparable by assigning weights, and once the weights exist the comparison is arithmetic again. Daniel Kahneman and Amos Tversky added the complication: people do not evaluate outcomes in absolute terms but against a reference point, and losses weigh more than equivalent gains. Any formula with a cost term and a value term is tilted before the arithmetic starts.

Deliberate installation has the evidence behind it. Peter Gollwitzer and colleagues, testing plans written in the if-then form, pooled ninety-four independent tests and reported a medium-to-large effect on goal attainment. Nearly twenty years later the same group came back with six hundred and forty-two tests and reported effects ranging from small to medium-large, depending on what outcome you measure, with larger effects when the plan used the if-then format, when the person already wanted the outcome, and when the plan had been rehearsed.

The second number is a range, not a figure. The often-quoted single decimal comes from the 2006 analysis and gets repeated as though it were a constant. It is not a constant. Neither analysis addresses publication bias in its abstract, and the earlier one sits in the period social psychology has spent the last decade re-examining. Forming if-then plans helps. The help is larger for behavior somebody has already committed to than for behavior they have not. Anyone quoting one number is quoting more confidence than exists.

The claim that the compiling actually finishes has much less behind it. The animal literature has a clean test for it, called outcome devaluation: train an animal to press a lever for food, destroy the value of that food, and see whether the pressing stops. An action still serving its goal stops. A compiled one keeps going. Christopher Adams found that the amount of training determines which you get, with moderate training staying sensitive and extended training going insensitive. Same animal, same lever, and the only thing that changed was repetition.

That is the cleanest operational definition of compiled I know of. In humans it does not survive contact with its own method. A 2026 paper argues that the standard human version of the test, which destroys a reward's value by feeding people until they are sick of it, frequently fails, because a participant who knows nobody will force them to eat the thing has not really stopped wanting it. When the devaluation fails, the participant keeps responding, and the experiment records that as habit. The authors report that habit-like responding concentrated in the minority for whom the procedure had not worked, and that with a better procedure most participants stayed sensitive to outcome.

Which moves the burden onto me. Anyone claiming a particular human behavior has gone fully automatic now owes an account of how they established that the goal stopped mattering.

The popular figure about how much of daily life runs on habit comes from two studies. Wendy Wood and her colleagues reported thirty-five percent in one and forty-three in the other, using a specific definition, on undergraduates, over one day and two days. Real numbers, narrow conditions, and a long way from a claim about people in general.

The question is what state a request is in. Three people currently answer it from judgment. The first thing that happens when you try to write the rule is that you discover the three of them do not agree, which nobody knew, because nothing had ever forced them to say their thresholds out loud.

So you start listing conditions. Has the customer been contacted. Has a part been ordered. Has a date been scheduled. Is somebody waiting on somebody else, and if so, on whom. Each of those has to become something a person can look at and answer without interpreting, and every one of them fights back. Contacted turns out to mean four different things. A voicemail counts for one person and not another. An email sent counts for one, an email answered for the next.

That is not a detour on the way to the formula. That is the formula being built. The conditional at the end is trivial once the terms are settled, and settling the terms is the entire job. Which is why the creed starts where it starts, with defining things, and why you cannot act upon that which has not been defined.

When it is done you have a rule that reads like nothing. If a part is on order and no date is scheduled, the state is waiting on parts. Anybody can run it. It produces the same answer on a Tuesday afternoon in March as it does at the end of a bad week, and the three people who used to disagree now disagree only about whether the rule is right, which is an argument worth having and a completely different one from the argument they were having before.

For the first two weeks after a rule like that ships, requests arrive that it answers cleanly and wrongly. A part is on order and a date is scheduled, so the rule returns one state, and everyone looking at it knows the real state is something else, because the scheduled date is three weeks out and the customer has not been told. The rule is not broken. It was written against a condition list that did not include how far away the date is.

So the list grows by one and the rule gets a second clause. The temptation at that moment is to keep going, letting every exception that surfaces become another branch until the rule is a decision tree forty nodes deep that exactly one person can maintain. Defining every variable to its deepest identifiable condition has a cost curve, and that is the shape of it.

What kept it from happening was counting. How many requests a month does this branch actually govern? When the answer was two, the branch did not get written. The case got routed to a person and the rule kept its shape. A formula that covers eighty percent of a queue and hands the rest to somebody who can think is a better object than one that covers ninety-eight and dies with its author.

The coat rule scales the same way and fails the same way. If it is cold, wear a coat, is unrunnable by two people. If it is below sixty degrees, wear a coat, is runnable by anyone, including somebody who thinks sixty is a ridiculous threshold. They can argue about the number now. The number is visible.

The boundary is where the conditions cannot be enumerated.

Friedrich Hayek made that case in 1945, about economies, and it generalizes without much effort. His claim is that the knowledge relevant to a decision never exists in concentrated form. It exists, in his words, as "the dispersed bits of incomplete and frequently contradictory knowledge which all the separate individuals possess." He gives that kind of knowledge a name: the knowledge of the particular circumstances of time and place. It is unorganized, it is not scientific, it cannot be aggregated upward into a central rule, and it is enormously valuable — but only, he says, if the decisions that depend on it are left to the person who holds it.

The technician on site knows that this customer has been patient twice already and is not going to be patient a third time. That knowledge exists nowhere in the conditions I enumerated. It is real, it is decision-relevant, and it is invisible to the formula. The rule will produce a correct answer and the wrong action.

So the domain of validity narrows. Formulas work where the conditions that matter are observable and enumerable. Where the decisive information is dispersed, local, and unwritten, the formula is not merely incomplete. It is systematically blind in a direction it cannot detect.

The blindness does not announce itself, and the people who can see it are usually the ones the rule is being written about rather than the ones writing it. The only instrument I have found that works is asking whoever currently makes the decision what they would need to know in order to make it, and then listening for the answers that cannot go on a form. When somebody says it depends on the customer, the follow-up is which customer and what about them, asked four or five times, until either a condition falls out or it becomes clear that none will. If none will, that is the answer, and it arrives before the rule is built instead of after it has produced a season of correct answers and wrong actions.

What a formula cannot do is report its own silence on a variable it was never given. It does not return an error. It returns an answer, with exactly the confidence it returns every other answer, and that confidence is a property of the arithmetic and not of the situation.

The blind spots run inward. Identifying your bias and then setting it aside, the method this whole book claims to use, says question everything, including your own conclusions. Compiling a formula says stop asking. Both cannot be running at full strength on the same rule at the same time. The best I have is that the questioning applies to the rule and the compiling applies to the execution, so that the rule gets audited on a schedule while nobody re-decides it mid-shift. That is a truce, not a resolution. The chapter on internal contradictions inherits it.

I also have no reliable way to tell a compiled formula from a dead one. A rule that still serves its purpose and a rule that outlived its purpose look identical from outside. Both run smoothly. Both produce consistent output. The devaluation test says to check whether the behavior changes when the goal stops mattering, and in a human organization nobody performs that experiment, because performing it means deliberately removing the reason for a process and seeing whether anyone notices.

Which leaves the honest position: every organization is running formulas whose purpose expired, and nobody can name which ones.

Hayek's dispersed knowledge is not the strongest case against the claim that almost everything in life can be reduced to a quantifiable formula.

Charles Goodhart, writing about British monetary policy in 1975, observed that "any observed statistical regularity will tend to collapse once pressure is placed upon it for control purposes." Marilyn Strathern later compressed it into the version everybody knows: when a measure becomes a target, it ceases to be a good measure.

Apply that to a rule of the form if X then Y. The moment a formula governs anything a person cares about, X stops being a neutral observation of the world and becomes a thing people produce. If waiting on parts is the state that stops the clock, parts get ordered. If contacted means an email was sent, emails get sent that nobody expects an answer to. Nobody has to be dishonest for this to happen. The rule created a gradient and water runs downhill.

The damage is worse than a gamed metric, because the formula was supposed to be the thing that made results consistent. Now it is the thing generating the behavior it was built to observe. The rule reads its own output and reports that everything is fine.

Goodhart's version is not about measurement corruption in the ordinary sense. It says the regularity itself collapses. The relationship you observed, the one you built the rule on, stops holding precisely because you built the rule on it. The formula destroys its own foundation by existing.

Every rule I have ever helped write was justified by an observed pattern. Requests in this condition tend to need that action. The pattern was real when I observed it. I then wrote a rule that acted on it, which changed what people did when the condition appeared, which changed the pattern. The evidence that justified the rule is evidence about a world that contained no such rule, and it stops describing the world the moment the rule ships.

There is no version of this I can engineer around by being careful. Careful is not the variable. A well-written rule with clear conditions and honest intentions produces the effect just as reliably as a sloppy one, and arguably faster, because people follow it.

And the failure is silent by construction. A rule that has corrupted its own trigger still runs. It still produces consistent output. It still passes every test I proposed earlier in this chapter, because two people with the same inputs still get the same answer. The consistency I offered as the benefit is intact while the thing consistency was supposed to deliver has quietly left.

Consistency is not the goal and never was. The goal was good outcomes, and consistency was a proxy I adopted because it is measurable and outcomes often are not. Goodhart's law is about exactly that substitution. I have spent a career installing proxies and calling it rigor.

I do not have a clean answer.

A formula holds where three things are true. The conditions it reads are observable without interpretation. The mapping from condition to action is stable over the period the rule runs. And the people the rule governs cannot manufacture its trigger, or gain nothing by doing so.

Goodhart's collapsing regularity forces the third condition, and it is testable. Before writing a rule, ask who benefits from the condition being true, and whether they can bring it about. If they can and they do, the rule is not a formula. It is an incentive wearing a formula's clothes.

Hayek's objection narrows it further. Where the decisive knowledge is local and unwritten, the formula can produce the answer but should not produce the action. The rule proposes; the person on site disposes, and the fact that they overrode it gets recorded, because an override rate is the only instrument I know of that measures how wrong a rule is from inside the system that runs it.

Run the three conditions against the status rule and it mostly holds. The conditions are observable, or were made observable at some cost. The mapping is stable across a quarter, which is longer than the rule has to run before anybody looks at it again. The third condition is where it gets uncomfortable, because a status that stops a clock is precisely the kind of trigger a person under pressure can manufacture, and no amount of definition fixes that. The fix is to stop letting that state stop the clock, which means the rule and the measurement have to be designed by people who are looking at each other.

Run them against the drive-thru on Route 8 and it holds all the way through. If A plus C is less than B, I get the double cheeseburger and fries. The conditions are observable to me, the mapping has not changed in years, and I am the only person the rule governs. Nobody profits by making me hungry. Goodhart has nothing to say about a formula with one participant who is also its author, and a great deal to say about every formula with two.

That is the dividing line the original was missing. Not the size of the decision, and not how consequential it is. Whether anybody other than the author has a reason to move the inputs.

None of that touches the coat. Temperature is observable, the mapping is stable, and nobody gains anything by making it cold. Small rules in stable conditions with no incentive gradient are where the claim holds without qualification: almost everything in life can be reduced to a quantifiable formula. Most of daily life is made of those.

What survives:

The original said almost and left the exceptions unnamed. Three of them now have names.

Two questions go forward unanswered. I still cannot tell a live formula from an expired one without running an experiment no organization will run, and until that is solved every claim about process drift rests on a gap. And the truce between compiling a rule and questioning everything is a scheduling trick rather than a reconciliation. Chapter 16 takes both of those on.