The thermodynamics of business data: decay, injection, and the economics of standing remediation#

G. Righter/ZnuLabs · Copyright © 2021-2026, Grover Righter#

And, like this insubstantial pageant faded, / Leave not a rack behind. — The Tempest, IV.i

Abstract. The data held in customer and resource platforms is not a static asset but a decaying one. We model a record’s correctness as a survival process with a finite half-life, and the platform’s aggregate quality as a reservoir subject to two opposing flows: continuous decay of standing records, and continuous injection of new records whose quality is bounded by the discipline of their capture. We show that such a system equilibrates at a quality strictly below that of its own inflow, that the fraction of jointly-usable records collapses geometrically in the number of fields a decision requires, and that one-time remediation relaxes back to equilibrium on the decay timescale. A standing remediation term is therefore not an operational preference but a structural necessity; and because the same exponent that governs the collapse governs the recovery, the return on that term is leveraged rather than linear. We close with the recoverability floor that bounds what any method can achieve.

Keywords: data quality, entropy, survival analysis, master-data management. MSC 60K25, 94A17. JEL C61, D83.

Motivation#

Every business platform—the CRM, the ERP, the marketing database—is treated by its owner as a store of value that persists until deliberately changed. This is the wrong physical intuition. The correct intuition is a warehouse of perishable stock. A contact’s employer changes; a title lapses; a company is acquired and its domain redirects; a phone number is reassigned. None of these events touches the record, which continues to display its now-false value with undiminished confidence. The asset degrades in place, silently, and the silence is the whole problem: there is no operational signal at the moment a true value becomes false.

To this steady internal decay we now add an unprecedented rate of external injection. Hand-keyed entry has always carried typographical and omission error. Acquired lists—trade-show scans, rented files—arrive pre-degraded. And most recently, automated meeting transcription deposits machine-inferred summaries of what an agent believed was said directly into the record of account, at volume, with review rates indistinguishable from zero. The platform is thus a reservoir filled by an increasingly turbid inflow while its standing contents rot. We make this precise.

Formulation#

Field-level decay#

Consider a single field of a single record, correct at the time of capture $t=0$. Let the event “the field’s value ceases to correspond to the world” occur with constant hazard $\lambda>0$. The probability that the field remains correct at time $t$ is then the survival function

The half-life $t_{1/2}$ is the interpretable quantity: the time by which a field of this class is as likely wrong as right. Distinct field classes carry distinct hazards—an email address bound to an employer decays far faster than a legal company name—so a record is a bundle of overlapping survival curves with heterogeneous $\lambda$.

THE FERRYMAN // Ο πορθμέας

Everything you know about your customers is dying.

Not fast. Not all at once. The way a good knife goes dull. The email that worked last spring. The title that was true. Someone moved. Someone quit. Nobody sent word.

The record sits there. Looks fine. It lies now.

That is the half-life. You have felt it for years. Here it is, finally, in a number you can point at.

The reservoir#

Let $Q(t)\in[0,1]$ denote the fraction of the platform’s records that are currently usable—the aggregate quality. Two flows act on $Q$. Standing correct records decay at hazard $\lambda$, removing $\lambda Q$ per unit time. Simultaneously the base turns over at relative rate $\mu$: a fraction $\mu$ of the reservoir is displaced per unit time by fresh inflow of quality $q$, the discipline of capture. Treating the reservoir as well-mixed,

Setting $dQ/dt = 0$ gives the equilibrium quality

This is the note’s first structural result. Since $\mu/(\lambda+\mu) < 1$ for any $\lambda>0$, the platform equilibrates strictly below the quality of its own inflow: $Q^{*} < q$. Perfect data capture ($q=1$) does not yield a perfect base; it yields $Q^{*} = \mu/(\lambda+\mu)$, the shortfall set entirely by the ratio of decay to churn. Decay guarantees that the destination is always beneath the door you pour through.

The modern injection regime is read directly off the equilibrium. Unreviewed automated transcription raises $\mu$ by orders of magnitude while lowering $q$. As $\mu \to \infty$, $Q^{*} \to q$: the equilibrium is dragged onto the inflow quality itself. A firehose of mediocre data does not merely add mediocre records—it pins the entire base at the mediocrity of the hose.

THE FERRYMAN // Ο πορθμέας

You cleaned it once. It felt good. It didn’t hold.

It never holds.

New data pours in. Trade-show lists. A temp’s typing. And now the machines, writing down what they think they heard in a meeting nobody read back.

The pool settles to the level of what flows in. Pour in mud, you get mud. Pour in a river of it, you get nothing else.

A clean database, held clean, was always a dream. Water finds its level. So does this.

The collapse of joint usability#

Per-field quality understates the damage, because business decisions are rarely made on one field. To place a record on a target grid—to say this is a mid-market manufacturer in the DACH region with a reachable buyer—a decision requires $k$ fields to be simultaneously correct. Let the per-field usable probability be $p$ and treat the field errors as approximately independent. The probability that a record is jointly usable for a $k$-field decision is

Usability collapses geometrically in $k$. The effect is severe at values of $p$ that feel comfortable in isolation. At a per-field quality of $p=0.80$—a figure a data owner would report without alarm—a four-field decision is usable for only $p^4 \approx 0.41$ of records; the remainder present as full but wrong. Table 1 tabulates the collapse.

Table: Joint usability $U=p^{k}$ against the number of required fields $k$, for three per-field qualities. The highlighted cell is the reference case of Section 3.

A database that is full can be substantially empty for any decision that matters. This resolves the recurring field observation that a segmentation report shows vast holes over a table reported as well-populated: each field was individually acceptable; their conjunction was not.¹

THE FERRYMAN // Ο πορθμέας

One field. Eighty percent. Sounds fine.

You need four fields to find a buyer. Industry. Size. Place. A way to reach them.

Do the multiplication. Forty percent.

More than half your list is gone. Not missing—worse. There, and wrong. Confident and wrong.

That is why the report comes back empty when the table looks full. Each part was pretty good. Together they were useless.

Remediation as a standing flow#

The reservoir model forecloses the intuitive intervention. A one-time cleaning sets $Q(0)$ high; the system then relaxes back toward $Q^{*}$ with time constant $(\lambda+\mu)^{-1}$. Quality bought once is quality lent, not owned—repaid to entropy on the decay timescale. Any durable solution must supply a continuous restoring flow.

Introduce a remediation term acting on the currently-bad fraction $(1-Q)$ at rate $\rho\ge 0$, representing a standing process that detects and repairs degraded records. The governing equation becomes

with equilibrium

Sufficient standing remediation drives quality arbitrarily close to unity. The remediation rate required to hold a target $Q_T$ follows by inversion:

Two properties matter operationally. The required rate is finite and modest for reasonable targets, scaling with the decay-plus-churn the system already sustains. And it diverges as $Q_T\to 1$: the last increment of perfection costs unboundedly, which is why the correct target is a held line, not a spotless base. The engineering object is a subscription to a rate, not the purchase of a state.

Definition (Recoverability floor). The repair term is not universal. A record is repairable only where a surviving anchor—a resolvable domain, an associated contact, any field carrying mutual information with the truth—permits reconstruction. Where the identifying signal was never captured, no $\rho$ recovers it; the mutual information is zero and the operation is not repair but reconstruction from an external source. Let $a\in[0,1]$ be the anchorable fraction. Effective remediation is $\rho_{\text{eff}}=\rho\,a$, and attainable quality is bounded by structure, not by effort.

THE FERRYMAN // Ο πορθμέας

So you don’t clean once. You keep a current running against the rot. Small. Always on. Enough to hold the line.

Not enough to reach perfect. Perfect costs everything and buys you almost nothing over good. You hold the line. That’s the job.

And what can’t be saved, you don’t chase.

A company name with nothing behind it—no site, no people, no thread back to the world—is not a customer. It’s a word someone typed. You don’t clean a word. You bury it, or you go buy the truth fresh. Then you move on.

The economics: why the answer is affordable#

The case for standing remediation is not merely that it works but that its return is leveraged by the very exponent that produced the harm. Let business value $V$ depend on the usable fraction, $V = V(U)$ with $U = p^{k}$. The sensitivity of usable population to per-field quality is

which is large precisely in the operating regime where $k$ is several. The exponent that drove the geometric collapse now multiplies the yield of every marginal point of quality recovered. The same structure, sign reversed: entropy compounds losses on the way down; remediation compounds recovery on the way up.

Against this stands the cost. Detection is a single pass over the base, $O(N)$: a diagnostic that returns the operating $Q$, the anchorable fraction $a$, and the field-conjunction $k$ the owner’s decisions actually require. This is cheap, and it is decisive—it converts an unpriced, invisible liability into three measured numbers, and it localises spend to the fraction where $\rho$ can act. Remediation is then a modest held rate $\rho_{\text{req}}$, applied only across the anchorable $a$. One compares this held rate against the leveraged value, not against the notional cost of an unattainable spotless base. On that comparison the standing process dominates both the do-nothing path, whose value bleeds as $Q\to Q^{*}$, and the periodic-cleaning path, whose quality traces a sawtooth whose mean sits well below $Q_T$.²

THE FERRYMAN // Ο πορθμέας

The exponent cut you on the way down. It pays you on the way up. Same arithmetic. You just flip the sign.

A little effort. Held steady. Aimed only where it lands.

Cheap to look. Cheaper than staying blind.

Nobody was ever ruined by finding out where they actually stood. They were ruined by not asking, and calling the silence good news.

Conclusion#

Business data is thermodynamic. It decays in place; it is diluted by an inflow whose quality it can never exceed; and for any decision of realistic dimension its usable fraction collapses geometrically. These are not defects of a particular platform or a lax operator—they are properties of the process, present wherever records meet time. It follows that quality cannot be bought as a state and kept. It can only be sustained as a flow: a standing remediation rate sufficient to hold a chosen line, applied across the fraction that carries recoverable signal, priced against a value that the collapse-exponent leverages in the owner’s favour. The diagnostic is the cheapest honest act available to a data owner—three numbers where there had been a comfortable silence. The remediation is the only durable one.

#