The Unpaid Fixes Are Already in Your Logs

At 09:14 a record is written with a wrong value. At 16:44 the correction lands. The seven and a half hours in between are a number the systems already keep and no dashboard displays.

Institutions run on unrecorded human repair. The repair is unpaid, the official record stays clean because somebody keeps it clean, and the whole position stands or falls on numbers like this one.1 A hospital’s measured efficiency, a claims system’s throughput, an EHR’s error rate: each is subsidized by work the system’s own metrics never see. That’s either an empirical statement or a mood. The difference is instrumentation, and most of the instrumentation already exists, much of it in records the institution keeps for its own purposes.

The Instruments Already Exist

Load at the moment of work comes first. NASA’s Task Load Index is a six-question form filled in about a task just completed: mental demand, physical demand, time pressure, performance, effort, frustration. It takes about two minutes. It was built and validated in aviation for exactly this purpose, as a subjective workload assessment that asks the operator how much load the task imposed. It asks what an engagement survey never does: how much of the person the task took, in that person’s own report.2 Filled in after a clinic session or a claims batch, it records demand where it lands instead of where the org chart says it should.

Second comes the distance between error and repair. Systems already timestamp every write. When someone corrects a wrong value, the two timestamps bracket the interval when the error was live. The measurement doesn’t require reading anyone’s screen.

Repair that lands outside paid time is the third. The documentation finished after the shift, the chart review on Saturday, the reconciliation done from the kitchen table. Measuring it needs two timestamps the systems already keep, one for when the work entered and one for when the worker’s paid presence ended. Nobody reads the content. The gap between the two moments is the number, and it doesn’t show up in any summary the software produces.

And the fourth is a reclassification rather than an instrument. The side spreadsheet, the paper note taped to the monitor, the personal group chat routing around a broken interface: dashboards file these as shadow IT and adoption friction. They’re outage reports filed by users, free of charge, describing the moment the official system stopped working. The hospital research is blunt about the price of ignoring them. Staff who fix the failure in front of them and move on are rewarded for resilience, and the organization often never hears what the failure was or how many more are coming.3

The last reading is already on file. A claim held in pending preserves the float, and the float draws yield for as long as it’s held. Every added layer of friction (the resubmission, the portal timeout, the multi-day review window) produces a measurable abandonment rate. The saving from abandonment gets booked whether or not anyone modeled it in advance. The institution suppresses claims without issuing a denial, and a denial is the only event that starts an appeal clock or triggers regulatory review. Both the float and the abandonment can be computed from the institution’s own aging report. Pending Is a No Nobody Signed names the state; the aging report supplies its accounting.

Why the Subsidy Goes Unmetered

The instruments are cheap. The obstacle is what a working meter says.

A clean dashboard is a sales asset. The throughput figure on the executive wall is throughput after subsidy: after the evening documentation, the workaround, the personal checklist. Meter the subsidy and the figure moves, which is a strong reason it goes unmetered. The number prices the exact gap the product was bought to close. A vendor’s incentive to conceal is the likeliest explanation and still a hypothesis: implementation cost, attribution, privacy and governance constraints can each keep a meter unbuilt too.

The same institutions aren’t shy about measurement elsewhere. Server uptime is instrumented to the millisecond, billing-code transitions are logged at the field level, and badge swipes are timestamped, because each maps to revenue or to a regulatory penalty. The badge reader knows to the second when the nurse left, and nothing has a field for the charting she then finished from bed. Whether anyone decided this matters less than it seems, because the incentives produce the same result either way. An institution that meters its workers’ compensatory labor takes on an obligation to pay for it the day the meter reads out. One that meters it while its competitors don’t prices itself out of the market. Omissions that carry the margin survive selection. The law has a name for a party that deliberately avoids learning what it suspects, and the doctrine asks about avoidance rather than malice. Whether it attaches to anyone in particular is for a court. What the record already shows is the pattern: every cost that maps to revenue is counted, and the one that maps to liability is not.

The structure is the one Appeals Test the Clerk, Never the Rule describes. A trial would obligate somebody to pay, so the trial doesn’t convene. A meter would obligate somebody to pay, so the meter doesn’t get built. Absence of measurement is then cited as absence of evidence, and the loop closes.

A Reading Gets Teeth in the Contract

A measurement changes nothing until it’s attached to money or authority, and the attachment point is the contract. The contract is also where the subsidy is written down. Enterprise software sells at premium recurring fees and is licensed as an informational aid or a drafting tool. So the hospital buys the system it runs on and is warranted a notepad. The vendor builds for the standardized path of a workflow and leaves the expensive edge cases to the buyer’s frontline staff, who do the reconciliation the license disclaimed, off the clock. The gap between what the marketing said the software automates and what the license refuses to warrant is the subsidy in the vendor’s own words. The off-clock hours that bridge it are already in the audit logs. Being Relied On Doesn’t Give a System Authority holds the general form: a fact about what a system can do never settles who answers for it.

Procurement is where the buyer has the most leverage, because the vendor wants the renewal. Four clauses carry most of the weight.

Acceptance runs on the buyer’s own data (the incomplete, contradictory, real extract) rather than on a demo set prepared for the sale. A screen that halts a task must name the person who owns the halt and the clock it resolves on. A status that only says pending is a defect with a price. Workaround logging counts as downtime: if staff keep a parallel spreadsheet to route around the product, those hours land in the availability figure the service credits are computed from. And every output that changes a record must trace to its source, so that correcting the source finds everything that relied on it.

The last clause is the one engineering waves through, so it deserves its number. Let repair time be how long a correction takes to reach every downstream system that consumed the error, and spread time how long the error took to reach them in the first place:

An incorrect value that propagates to forty systems in an afternoon and takes three weeks to retract has a D in the hundreds. A pipeline with no measured D shouldn’t get to claim a number at all. D is an ordinary reliability quantity. Its absence from AI procurement is the tell.

None of this needs a theory to say out loud. In an architecture review the whole position reduces to three sentences: this feature offloads forty minutes of manual reconciliation onto the triage desk, which is where it fails to scale; this model has capability without an authorization record; this workflow has no rollback graph, so its repair latency is unbounded. All three are checkable, all three carry a price, and none of them requires anyone to agree about justice first.

Which Meters Stay Pointed Upstream

Of the standing objections, surveillance is the serious one, and Measure the Machine, Not the Worker works out which of these instruments stay pointed upstream at the design and which turn into a whip.

Read against each other, the contracts, the logs and the aging report prove where the operating margin comes from. The proof never has to establish that anyone involved is evil. The practices read as commercial and legal choices, which is what they are. A choice can be defended only on the grounds this reading names: who bears the liability, who does the labor, and who collects the profit.

The diagnosis has been available for some time. The instruments are older than it is. What the argument lacked was never a way to see the subsidy but a place to stand while demanding it be paid for. A timestamp is a place to stand, and so is a contract clause. Metered honestly, the subsidy becomes an invoice. An invoice is only as good as the claim behind it, and the claim needs three things the measurement alone can’t supply: an enforceable form, a claimant with the standing to bring it, and consequences for nonpayment (the due date, and what happens when it passes). Before an Automated System Goes Live, Post a Bond shows what a bond does to the last of those.

The reader who is asked to approve the next system (the renewal, the contract clause, the demo) has the three requirements already reduced to one question with a due date: who owns the halt, and what happens when the clock passes? The clause that answers it is the invoice in enforceable form. The clause that can’t is the subsidy still running unread.