What Counts

I had reached the point where the evidence itself needed an explanation.

An agreement had survived. A hazard was being managed. A permission remained in a document. Each fact established something about the inheritance I had been examining. None, alone, established that people would receive the conditions and powers needed to judge what they inherited. Yet a present institution still had to act. It needed an account of those conditions before the future could supply its own.

I followed that need beyond the question of inheritance. The promises in my catalog acquired practical consequences through things more manageable than the lives they concerned. A claim entered a category. A condition received a score. A recorded action became a reason for another action. These were among the means by which an obligation reached anyone at all. Their usefulness was part of the difficulty.

An institution unable to distinguish relevant circumstances would have little basis for consistent treatment. An institution unwilling to simplify would struggle to act beyond the immediate knowledge of its members. I admired instruments that made distant needs perceptible. The admiration gave me a reason to examine how far their findings reached, rather than to dismiss them for having limits.

The question was no longer whether an account left something out. Every account did. I wanted to know what was allowed to follow from what remained in it. A representation acquired a different significance when it ceased to inform a decision and became sufficient to settle one. The information might stay the same while the authority attached to it changed.

That attachment crossed several of my provisional entries. A model standing in for its target and a metric standing in for a condition were not identical operations. Nor was either the same as rewarding a report that looked favorable. Before deciding whether an instrument had displaced its purpose, I needed to follow the work it had been assigned.

Enough to decide

A signal that gives someone reason to investigate does not necessarily give that person reason to impose a consequence. Investigation may be justified precisely because the signal leaves something important unresolved. If the next action treats the uncertainty as settled, the account has done more work than the evidence supplied.

I already knew that a model could perform competently at the wrong task. The further question concerned the institution receiving its result. What had it agreed to regard as enough? That question applied to a prediction, but it also applied to an ordinary file. No elaborate computation was required for information to acquire a consequence beyond its warrant.

The distinction became particularly clear when I followed evidence that was already available. In a United States review of Medicare Advantage denials, the inspector general identified additional clinical criteria, overlooked documentation, and manual or system errors as different reasons that requests meeting the relevant rules had nevertheless been denied. One authorization case contained supporting information already supplied with the request. The organization later reported reversing its decision after the inspector general sought information about it.¹

The patient had not needed the world to become more measurable. The relevant information had entered the account. What failed was its use in reaching the decision. That located the difficulty differently from a missing observation or an inaccurate prediction. A demand for more data, made without that distinction, risked asking someone to supply again what the institution already possessed.

I also had to keep the reversal inside the finding. The decision did change. An account of evidence permanently denied all corrective force would have misdescribed this record. But the reversal established a change in authorization, not what treatment the person eventually received or what happened to their health. The correction had a reach, and the record did not permit me to extend it further.

This was a less elegant classification than a closed institution confronting an excluded reality. It was also more useful. It distinguished the information supplied, the judgment made from it, and the intervention that changed that judgment. Those differences identified where correction had operated. Calling the whole sequence opacity would have hidden some of what the investigation made visible.

The existence of a standard did not explain the failure by itself. Standards can protect people from arbitrary distinctions and from decisions based on insufficient support. A request does not become warranted simply because refusing it causes difficulty. The relevant comparison was between what the decision required and how the supplied evidence met that requirement. I could examine that relation without assuming either the claimant or the institution must always be right.

I became more attentive to the word insufficient. It sometimes describes the evidence. It can also describe a conclusion reached about evidence, and that conclusion is open to error. Confusing the two makes the judgment disappear into the material being judged. The file seems to lack something because the institution has said that it does.

A threshold adds another decision to this process. It distinguishes what happens on one side from what happens on the other. That distinction needs a justification suited to the action it governs. A boundary useful for assembling a research population cannot establish, merely by existing, whom a service ought to exclude. The precision of a boundary does not supply the reason for placing it there.

Nor does the boundary alone reveal the distribution of its burdens. For that I would need to know who encountered it, which differences mattered, and what treatment followed. My catalog made it easy to recognize the possibility of a transfer of risk. Recognition did not relieve me of discovering whether that transfer had occurred in the arrangement before me.

The same restraint applied when I moved from measures to models. A model has a target, assumptions about how its inputs bear on that target, and circumstances in which its performance has been examined. The institution adds a use. That use determines which errors matter and what follows from them. An error tolerable in deciding where to look may be intolerable as the sole basis for deciding what someone owes.

I could now see why an instrument's limits belonged in the account of its authority. A caveat filed beside the method does little if the decision proceeds as though the caveat had been resolved. Respecting a limit means allowing it to affect what the institution concludes or does. The qualification must reach the action, not merely survive in the documentation.

This did not make judgment an alternative to measurement. It made judgment part of the work measurement required. Someone still had to decide what the account warranted, and that decision needed reasons. A quantified result could sharpen those reasons without relieving anyone of giving them.

The distinction left me with a further difficulty. Institutions did not only use accounts to decide what should happen. They used changes in those accounts to judge what had happened. A better-looking result seemed to promise a simpler conclusion. I wanted to know what had moved before accepting the direction of movement as its meaning.

When the number changes

A number compared with its earlier value carries an invitation to read time into it. Something has increased or diminished, and the difference appears to describe what happened between the two observations. But that reading depends on how the observations were made. A change in the world and a change in the account of the world can occupy the same difference.

Maintenance records brought this difficulty back to the question of inheritance. In its 2023 review of four United States civilian agencies, the Government Accountability Office found that estimation changes and assessment cycles helped explain increases in reported deferred maintenance and repair. Some methodological changes likely improved the estimates. Their contribution to the increases was not adequately explained in the agencies' reporting.²

The unfavorable number therefore deserved more attention than an unfavorable verdict. A better assessment might reveal obligations an earlier estimate had missed. That would worsen the account of the institution's position while improving its knowledge of the work ahead. Actual deterioration remained possible; the two explanations were not mutually exclusive. The reported increase did not divide itself between them.

I resisted the convenience of calling this statistical camouflage. Missing explanation mattered, but a method that detected more of a problem was not thereby making the problem disappear. If I classified every altered estimate as evasion, I would reward the preservation of ignorance whenever correcting it made the report less flattering.

The account needed a history of its own. When had the relevant assets been assessed? Had the method changed? What did the estimate now include? These questions concerned how the result was produced. Without them, a reader was being asked to interpret a movement while part of the movement's cause remained inside the measuring practice.

The distinction changed what improvement in the report should mean. Making the estimate more interpretable was useful even if it did not repair a building. In its subsequent follow-up, GAO recorded added information about the mission importance of the Department of Energy's backlog as an implemented recommendation. It also recorded an evaluation that did not find expanded modeling financially advantageous.³

I liked the discipline of that second result. Examination had not become an obligation to purchase a more elaborate answer. A model was still a means whose value required judgment. To decline one after evaluating its usefulness was not, by itself, a retreat from knowledge.

There was a practical consequence to keeping these achievements separate. A clearer account could improve the basis for deciding what to repair while leaving the repair undone. That distinction did not reduce reporting to ceremony. It identified the next task. The error would be to let completion of the explanation stand as completion of the work it explained.

The direction of a reported change was therefore insufficient to identify improvement. Even the word improvement needed an object. The estimate might improve in accuracy, the decision in responsiveness, or the condition in which people lived. Those changes could support one another, but an account of one did not establish the rest.

I wanted an observation made through another route. The Office for National Statistics offered such a comparison in England and Wales. Its 2024 analysis linked rising police-recorded crime between 2015 and 2019 chiefly to improved recording, while the comparable crime-survey series generally fell. Removing the violence categories most affected by recording improvements brought the trends closer together. The disagreement itself helped locate a change in how the record was made.⁴

Yet the survey did not become an unrestricted account of crime by being separate from police records. It covered people living in households; police records also included offences against businesses and people outside that population. The periods covered differed too.⁵ I needed to know where the accounts could answer the same question before asking one to settle the other. Otherwise I would correct an overextended inference by giving its replacement the same excessive reach.

Incentives introduced a different reason for examining the movement. If a result carries a reward, those whose work produces it have a reason to attend to what is rewarded. That need not be a defect: the reward may direct effort toward the intended condition. The question is whether the condition and the report remain connected when people respond.

The audit experiment I had examined earlier, in which the plants paid their own auditors, took place in Gujarat. It remained important because it did not leave that question inside the audit reports. Changes to auditor assignment, payment, and checking improved reporting, while separate endline sampling found lower pollution. The intervention was a package; the result did not identify one component as the entire cause. It did show why evidence of improved reporting and evidence of an improved condition deserved separate examination.⁶

I had no reason to turn that achievement into another proof that institutions only repaired their appearance. The possibility of distortion had not made truthful measurement impossible. Changing the arrangement under which the account was produced helped it serve the condition more faithfully. The repair altered what I could responsibly say about the instrument.

The audit experiment gave the separate observation a defined task: it examined the condition the reform was intended to improve. That connection was what I had needed when comparing the crime accounts. Independence alone did not supply it. The maintenance estimate still required its own history, and the experiment did not divide that backlog into better observation and actual deterioration. Keeping the questions distinct let the evidence from each do its proper work.

Truth did not become less important because records needed interpretation. It became more demanding. A truthful statement about an estimate remained a statement about an estimate. To use it as a statement about the condition required a defensible connection between them. The authority of the number could not supply the missing connection merely because the arithmetic was correct.

Knowledge involved more than possessing the number. It included understanding what had been observed, how the result had been produced, and where the inference should stop. That understanding did not always call for a longer report. Sometimes it required a short explanation that changed the meaning of the result already displayed.

I began to read claims of correction with the same attention. A corrected figure was a changed account. A reversed decision was a changed action. Each might matter greatly, but neither told me everything that had happened to the people or conditions involved. The word corrected was no more entitled than the word improved to leave its object unspecified.

The reach of correction

The reversal in the authorization record returned to me with a different significance. It showed that relevant information could become consequential through reconsideration. It did not show that the original delay had vanished. I needed language in which the correction retained its value without being made to accomplish something time had already placed beyond its reach.

That was not a reason to demand the impossible before recognizing repair. An institution that changes a wrong decision has done something different from one that leaves it in place. The remaining inquiry concerns what the correction restored, what consequences persisted, and what the available record can establish about either. An imperfect recovery is not an imaginary one.

I found a related distinction where the account concerned a person's authorization. Here the record did more than describe a condition to be evaluated. It was used to support an entitlement to act. The transition from an event to permission deserved particular care, because a record could accurately preserve what someone had done while leaving open what they had authorized.

The Federal Trade Commission's complaint against Epic Games alleged such a gap in billing. It described charges without adequate authorization, purchase controls that produced unintended transactions, and account deactivation after consumers disputed charges through their payment providers. That last action also denied access to content bought earlier. These were regulatory allegations resolved through a consent order, not admissions by Epic or findings after a trial.⁷

What interested me was the relation between the recorded action and the authority claimed from it. A purchase control is meant to turn a choice into a transaction. The existence of a transaction cannot, on its own, establish that the control obtained the choice it was meant to carry. Otherwise the consequence would become its own evidence of authorization.

The objection made the distinction harder to ignore. In the Commission's account, disputing a charge and losing access to previously acquired content connected the contested transaction to a larger consequence. I wanted to know what permission the original action supplied and what justified the response when that permission was disputed.

Convenience did not answer the first question, though it was a legitimate reason to make repeated transactions easier. Fraud prevention did not answer the second merely by naming a legitimate concern. Both functions needed to be related to the particular scope of action. A reason for having a procedure was not yet a reason for every consequence attached to it.

The final order required informed authorization for charges and a simple means of withdrawing authorization for future billing. It prohibited denial of account access for disputing a charge. It allowed adequately bounded future authorization rather than treating every continuing permission as invalid.⁸

That boundary mattered to my understanding of consent. People need ways to authorize continuing arrangements without repeating every earlier decision. The question is what the permission covers and how it remains answerable to the person who gave it. A recorded act does not resolve those questions simply by being definite. The precision of a timestamp is not the precision of an agreement.

I kept the remedy's requirements separate from evidence of their effects. In June 2025 the FTC reported issuing nearly a million refund payments totaling more than $126 million. This was administered redress, beyond the announcement of an order. It did not establish that every payment had been received, every affected person reached, or the redesigned practices durably evaluated.⁹

Money returned addressed a loss. It did not make the earlier action authorized. A rule governing future billing addressed another part of the arrangement, while evidence about how that rule worked would answer a further question. The remedy had begun to change the relationship; its eventual reach still required examination.

The consent comparison changed the way I held the earlier records together. In the denial review, supplied information had failed to produce the decision it supported. In the billing allegations, an action was treated as supporting authority whose scope was disputed. Both required attention to the meaning attached to a record, but they were not the same failure. Improving the collection of facts alone would not settle either relationship.

The distinctions also prevented success from acquiring an immunity of its own. A corrected decision did not certify a whole procedure. A useful report did not settle the condition of everything it described. Effective redress did not establish that every future authorization would be adequate. Success deserved an account of its reach for the same reason failure did: other judgments would be made from it.

The work did not require abandoning the instruments. I had seen reasons to value a clearer estimate, a reconsidered decision, and a measurement arrangement that improved the condition it examined. Their limits made their achievements more exact. They also kept those achievements from deciding questions the evidence had not reached.

One such question now lay beyond the boundary of each account. Even where I could follow what a decision warranted and what its correction changed, I had not yet established what happened when its consequences entered the circumstances of another decision. Separate accounts did not become an account of their interaction merely by being accurate. I needed to follow the decisions beyond the limits within which each had been examined, to the conditions they helped make together.