The Error That Balances

I.The First Known Thing — the check that is one number wide

accounting · the trial balance, the suspense account, and the errors neither can raise

A ledger kept by double entry records every transaction twice, once as a debit and once as a credit, so that the sum of all debits must equal the sum of all credits. The check on the whole of the bookkeeping is the trial balance: the two totals are extracted and compared. It is the oldest routine audit in commerce and it has a property worth stating plainly. However many accounts the ledger holds, and however many thousands of entries have been posted to them, the check is a comparison of two numbers. It has one degree of freedom against the whole of the work it examines.

where the difference is parked

When the two totals disagree and the cause is not at once apparent, the bookkeeper does not leave the accounts unprepared. The difference is posted to a suspense account, a temporary holding account opened for amounts whose proper destination is not yet known. It is a parking space and never a home. The suspense account must be cleared to nil before the final accounts are drawn up, each underlying error corrected by its own journal entry, and it must never appear in the statement of profit or loss or in the statement of financial position. A set of published accounts therefore contains no suspense account, whatever happened during the year.

the six errors the totals cannot see

What disturbs the two totals is a one-sided error: an entry posted on one side and not the other, a partial omission, a figure transposed in one account alone, an error in the additions. These are the errors the suspense account receives. The bookkeeping syllabus then lists, as a separate category to be learnt, the errors the trial balance does not reveal at all. A transaction omitted altogether, both sides missing, leaves the totals equal. An error of commission — the right amount on the right side, in the wrong account of the right class, a receipt credited to the wrong customer — leaves them equal. An error of principle, a payment for a machine charged to repairs rather than to capital, leaves them equal. An error of original entry, where the same wrong figure is posted to both sides, leaves them equal. A complete reversal, debit and credit exchanged, leaves them equal. And compensating errors, two or more unrelated mistakes whose effects on the two totals happen to be equal and opposite, leave them equal. In each of these the ledger balances and the ledger is wrong, by an amount that is not bounded by anything the check can report.

what the parked difference can and cannot report

One consequence is seldom remarked, though it follows at once. The suspense balance is drawn entirely from the first class. It is composed of one-sided errors and nothing else, because nothing else reaches it. Its size is therefore a measurement of how much detectable error occurred and carries no information, in either direction, about how much of the other kind did. A year that ended with a large suspense difference and a year that balanced at the first attempt are equally silent about omissions, principles, reversals and compensations. And because the account is cleared before the books are closed, the finished record does not report even the detectable share. The only way to know what the trial balance cannot see is to reason about the structure of the check, in advance, from the way the totals are formed.

II.The Second Known Thing — the readouts of a safety study

drug discovery · toxicity and safety pharmacology

A repeat-dose toxicity study gives a compound to groups of animals at ascending doses for a fixed period and then takes a defined set of measurements. Body weight and food consumption. Clinical observations. A haematology panel and a clinical chemistry panel. Organ weights at necropsy. Macroscopic examination, and histopathology of a long prescribed list of tissues. From the pattern of what moves and at which dose, a no-observed-adverse-effect level is set, and the margin between that level and the intended human exposure is the central quantity that the whole programme carries forward.

each readout is a sum

Almost every one of those measurements is an aggregate, and the arithmetic matters. An organ weight is the sum of the masses of every cell population in the organ together with its blood and its water, so hypertrophy of one compartment and loss of another enter it with opposite signs. A serum analyte is a concentration held at the balance of its appearance and its removal, so a change in production and a change in clearance enter it with opposite signs. A blood count is the balance of production against destruction, and enters them the same way. A body weight is the sum of everything. In each case the quantity reported is a scalar standing in front of a space of changes with many dimensions, and the readout is blind, exactly, to every change whose signed contributions to it come to nothing.

three cases the field has documented

That blindness is not hypothetical, and three instances of it are well established. Serum creatinine rises only with marked damage to functioning nephrons and is unsuitable for detecting early loss of filtration: creatinine is produced at a rate set by muscle mass and removed by glomerular filtration together with proximal tubular secretion, and as filtration falls the secreted fraction becomes a larger share of the total cleared, so the concentration is held while the filtration behind it is going. In compensated haemolysis the marrow raises red cell production enough to keep haemoglobin normal while red cell survival is shortened, the compensatory capacity of erythropoiesis being conventionally put at several-fold, so the count is normal and the destruction is ongoing. And an organ weight can be unchanged where one compartment has enlarged by the mass another has lost, which is arithmetic rather than physiology and needs no mechanism at all.

how the field holds them

These cases are known, and the way they are held is worth noticing, because it is organ by organ. The renal case is discussed as the insensitivity of creatinine and as renal functional reserve, a property of the kidney. The haematological case is discussed as compensated haemolysis, a property of the marrow. The adaptive-versus-adverse literature concerns how a change that has been detected should be classified. The biomarker programmes seek more sensitive markers of the same quantities. Each of these is a fact about a tissue, found when it was found, and the remedy proposed in each case is a better instrument for that tissue. What is not held anywhere is the general form: that the set of changes a readout cannot see is determined by how the readout aggregates, that it is the same set for every compound, and that it is computable before any compound is given.

III.The Bridge

no paper on the trial balance cites a paper on toxicological readouts; each hall examines a system by requiring sums to come out right, and neither has written down what that costs

The halls do not meet. The literature of the trial balance is the bookkeeping syllabus and the audit textbook; the literature of nonclinical readouts is regulatory toxicology and clinical pathology. Neither cites the other and no review spans them. Four neighbours must be set aside, and they are nearer here than usual. This is not functional reserve, which is the same phenomenon held as a property of an organ, discovered separately in each one, where the claim below is about the arithmetic of the readout and holds for organs with no reserve at all. It is not endpoint sensitivity, nor assay precision, which move the smallest change a readout can resolve: precision moves a floor, and cancellation is not beneath a floor but exactly at nil, where no precision reaches it. It is not the adaptive-versus-adverse question, which governs the classification of changes that were seen. And it is not the biomarker programme, which seeks a finer measure of the same aggregated quantity, where what is wanted here is a decomposition of it into signed parts.

X: an examination conducted through aggregates can raise only what moves one, so its blind set is fixed by the arithmetic of the sums and not by the quality of any measurement

Call it cancellation blindness. Where a system is examined by taking a set of aggregates and asking whether they have moved, the examination can raise only those changes that move one, and any combination of changes whose signed contributions to every aggregate taken come to nothing passes undetected whatever its magnitude; the reach of the examination is therefore fixed by the number and the arithmetic of the aggregates set against the dimension of the changes possible beneath them, and not by the precision, the replication or the duration of any measurement. Three things follow, and they are not one thing said three times. The first is that the residual the examination produces — the difference it cannot account for, the finding it cannot attribute — is drawn wholly from the detectable class, so its size reports the detectable share of the error and is uninformative in either direction about the cancelling share, which makes a clean examination and a troubled one equally silent about the blind set, and makes the blind set computable only from the structure of the aggregates in advance. The second is that the blind set has an orientation, and where the system under examination contains machinery whose function is to hold a quantity constant, and the examination measures that quantity, the blind directions are precisely the ones that machinery produces: the examination is then least able to see the changes the system is working hardest to offset, and sees them worse the better the regulation is. In a ledger this alignment would be a coincidence; in a body it is the design, and it is the reason the blindness is not a rare accident but a standing condition of the better-regulated organs. The third is that the residual is cleared before the record closes — the suspense returned to nil, every finding in a study attributed before the report is signed — so the finished record carries no measure of its own blind set, and nothing in it can be read as evidence that the set was small.

IV.The Question That Becomes Possible

neither discipline would ask this alone

The accountant has no reason to ask it. The errors the trial balance does not reveal are a list to be learnt and then guarded against by other controls — reconciliation, sampling, review — and no one computes the size of the space those errors inhabit, because a ledger has no interest in its own blind subspace once a second control covers it. The toxicologist has no reason to ask it either. Each known cancellation is a fact about a tissue, met when it was met, and the remedy is a better marker for that tissue; the question of whether the cancellations across all the readouts have a common form, computable in advance from the way each readout sums, is not a question the organ-by-organ framing generates. Held together they give a quantity that can be computed from study data already collected.

the compensation margin hypothesis

For a given organ and a given standard readout, does a compound have a compensation margin — the ratio of the lowest dose at which an effect becomes detectable in the aggregate readout to the lowest dose at which it becomes detectable in a readout decomposing that same aggregate into its signed components — and is that margin governed chiefly by the aggregation structure of the readout and the tightness with which the measured quantity is regulated, rather than by the compound or its mechanism, so that it is shared across compounds acting on the same organ through different routes and can be stated before any of them is given?

why it would matter

Because it would give every standard readout a second number beside its detection limit: not the smallest change it can resolve but the set of changes it cannot resolve at any size, with a dose ratio attached. Because it would tell apart two things that a no-observed-adverse-effect level reports identically — a dose at which nothing was happening, and a dose at which two things were happening that summed to nothing — and the two warrant different conduct, the first a margin and the second a decomposing measurement or a stated declaration that the organ is examined only through a quantity it defends. And because it would say whether the known cancellations are a handful of physiological curiosities, as they are now held, or instances of a property that every aggregated readout has and that no study reports.

V.How It Could Be Proven — or Broken

a candidate must be breakable, or it is merely poetry

An intercognix candidate earns nothing until it names its own test. This one needs no new animal and no new compound at the outset. What it needs is studies in which an aggregate readout and a decomposition of the same aggregate were taken on the same animals at the same doses, and those exist in numbers: kidney work pairing serum creatinine with measured filtration and segment-specific injury markers and histopathology; haematology pairing haemoglobin with reticulocyte counts and red cell survival; hepatic work pairing organ weight with compartment-specific counts.

the honest status

Note the narrowness. The claim is about readouts and not about compounds. It does not say that any approved drug is unsafe, nor that nonclinical studies miss harms at any particular rate, nor that any margin now in use is wrong. It says nothing about clinical monitoring, nothing about efficacy, and nothing about which readouts ought to be adopted or retired. It does not assert that a decomposing measurement is available for every aggregate, and for several of them none is. The whole of the claim is that the set of changes an aggregated readout cannot see is determined by how it sums and not by how well it measures, that this set is orientated towards the quantities the body holds steady, that the resulting dose ratio is a measurable number, and that it is at present treated as a short list of organ-specific curiosities rather than as a property of the measurement itself.