The Sound That Never Changed a Word

I.The First Known Thing — the difference that does no work

phonology · a sound counts only where it can change a meaning

A language has far more sounds in it than it has units. The p of pin is released with a puff of breath and the p of spin is not; a spectrogram separates them without difficulty, and a speaker of Hindi or Thai hears them as two consonants that distinguish words. An English speaker hears one. Phonology, as the Prague school framed it in the nineteen-thirties and Trubetzkoy set out in the Grundzüge der Phonologie, explains why without appealing to the size of the acoustic difference at all. A distinction between two sounds is a unit of the language — a phonemic opposition — only if there exists an environment in which the two can be exchanged and the exchange alters what is meant. The test is the minimal pair: pin and bin differ in one sound and in meaning, so p and b contrast. No such pair exists for the two kinds of p, and so however different they sound they are variants of one unit, allophones, and the difference between them belongs to the speaker's mouth rather than to the language.

complementary distribution: two sounds that never meet cannot contrast

The reason no minimal pair exists for the two p sounds is structural. The aspirated one appears at the start of a stressed syllable and the unaspirated one after s; each occurs only where the other does not. Sounds arranged this way are said to be in complementary distribution, and the phonologist's rule is firm: two sounds that never occupy the same environment have never been given the chance to distinguish anything, so their distinctness cannot be established, and they are assigned to one phoneme by default. This is not a claim that the sounds are similar. It is a claim that the question of their distinctness is undecidable from the language's own evidence, because the language has never placed them in a position where the answer could show.

neutralisation: an opposition that holds in one position and lapses in another

Between full contrast and no contrast there is a third condition Trubetzkoy took particular care over. In German, d and t contrast in the middle of a word, yet at the end of a word every voiced consonant is pronounced voiceless, so Rad, a wheel, and Rat, advice, are pronounced identically. The opposition has not been lost from the language; it has been neutralised in one position. Trubetzkoy's device for this was the archiphoneme, the unit that is neither d nor t but the pair with their difference suspended, and it is the operative unit wherever the neutralising environment applies. The distinction is therefore not a property of the sounds but of the sounds together with an environment, and the same two sounds can be two units in one position and one unit in another.

functional load: the work an opposition does predicts whether it survives

Oppositions are not equal in how much they do. Some distinguish hundreds of word pairs and some distinguish a handful, and the phonologist's name for this quantity is functional load, a term Mathesius introduced and Martinet made into a hypothesis about change: an opposition that carries little load is the one a language can afford to lose, so low-load contrasts should merge more readily over time. The claim spent decades contested. Hockett proposed a way to quantify it; King, testing it against the history of the Germanic languages in 1967, found no support and said so plainly; Surendran and Niyogi later gave it an information-theoretic form. Then Wedel, Kaplan and Jackson, in 2013, counted minimal pairs across the histories of several languages and found that the number of word pairs an opposition distinguished did predict whether it later merged. The result is a mechanism. What kept a contrast alive was the work it did — the count of environments in which it changed meaning — and not the phonetic distance between its members, which can be large for an allophone and slight for a phoneme.

II.The Second Known Thing — the subtype that may or may not be an indication

drug discovery · disease definition and the choice of indication

Before a programme has a target, it has a disease, and before it has a disease it has a decision about where one disease ends and the next begins. Classifications of disease are revised on a schedule: the World Health Organization's classification of blood and lymphoid tumours has appeared in successive editions across two decades, its classification of brain tumours has been redrawn around molecular markers, and consensus criteria in rheumatology, respiratory medicine and psychiatry are reissued with subtypes added, merged and renamed. Each edition is a set of oppositions — this is a separate entity, that is a variant of another — and a programme choosing an indication chooses from among them. The choice fixes the eligible population, the comparator, the endpoint and the label, and a mistaken split or a mistaken merger is paid for in a trial that enrols people who cannot respond, or excludes people who could.

how the splits are made: by difference

The grounds on which classifications split are, overwhelmingly, grounds of difference. Two groups of patients are separated because their tumours carry different fusions, their cells express different markers, their transcriptomes cluster apart, their courses differ, their histology looks different under the microscope. The field's shorthand — lumping and splitting — has an active literature in genomics, in sepsis and in the inflammatory diseases, and the recurrent question is how much difference justifies a split. Molecular resolution has made this harder rather than easier, because every cohort now yields clusters: adult diabetes has been proposed to fall into five, asthma into a growing set of endotypes, and the argument in each case turns on whether the clusters are real, which is asked as a question about the size and stability of the difference between them.

what actually decides whether a split holds

Yet the splits that have proved most durable are not the largest. The distinctions that organise breast cancer — hormone-receptor status, HER2 status — were made real by endocrine therapy and by trastuzumab, treatments that work on one side of the line and not on the other; the histological division of non-small-cell lung cancer into squamous and non-squamous forms was of small consequence to treatment for decades and became a hard boundary when a chemotherapy proved effective in one and not the other, and a targeted agent proved dangerous in one and not the other. Conversely, entities separated on grounds of difference have lapsed from classifications when nothing in practice ever told them apart. The field knows this in the phrase clinical utility and in the concept of a predictive biomarker, but it treats them as properties a split may happen to acquire, and it does not use them as the definition of what a split is.

and a class of splits that no trial has ever tested

There is, besides, a class of distinctions that no treatment has been positioned to test at all. Paediatric and adult forms of a disease are separated by age and studied in separate trials; sex-specific forms are separated by sex; the same tumour in different organs is treated by different specialties and enrolled in different studies. Whether the members of such pairs respond differently to the same treatment is not known to be true or false; it has never been asked in a way that could answer, because the two are never placed together in the one environment where an answer could show. The classification records them as distinct, on grounds of difference, and the record cannot be checked.

III.The Bridge

no phonology paper names a tumour

The halls share nothing. Phonology belongs to linguistics and the study of sound change; disease classification to pathology, epidemiology and the regulatory craft of defining an indication. No review spans them, and the familiar observation that classification is hard predicts nothing and is not the bridge. Nor is the bridge the trivial remark that a subtype should be actionable, which the clinical hall already makes. Two neighbours must also be set aside. It is not the crossover spread of an earlier specimen, in which a threshold drawn through a continuum sits differently for different observables; there the question is where a line falls, here it is whether a line exists. And it is not the exclusion-defined indication of another, whose boundary is borrowed from its neighbours; here the boundary is drawn positively and the question is whether it does anything. What the two halls share is narrower and structural: a system of categories acted on by a set of operations, in which the distinctness of two categories is not a property of the categories but of the operations, and is defined by whether any operation returns a different result when one is substituted for the other.

X: distinctness lives in the operations, and is counted, not measured

Call it contrastive load. Where a set of categories exists to be acted on — words to be told apart, patients to be treated — two categories are distinct only if there is an environment in which exchanging one for the other changes the result of the action; where no such environment exists the difference between them, however large and however real, does no work and the categories are variants of one unit. Four things follow. Distinctness is a property of the set of environments and not of the magnitude of the difference, so a large difference can be allophonic and a slight one phonemic. Two categories that never share an environment cannot contrast, and their distinctness is undecidable rather than unknown. An opposition can hold in some environments and be suspended in others, and in the suspending environments the operative unit is the pair with its difference removed. And the number of environments in which an opposition changes the result — its load — predicts whether it survives revision, independently of the difference between its members. Read into the clinical hall, each has a consequence the hall does not currently state. A subtype distinction is not made real by the molecular distance between its members but by the existence of a treatment environment in which the substitution changes the outcome, and until one exists the split is provisional however well it clusters. Paediatric and adult, male and female, one organ and another are, for a distinction never tested across them, in complementary distribution: the classification's verdict on them is a default and not a finding, and should be recorded as such. An opposition that is real under one drug class and absent under another makes the indication drug-class relative — the archiphenotype, the union of subtypes with their difference suspended, is the correct indication for a drug that does not see the difference, and enrolling by the finer split there buys nothing but a smaller trial. And the count of treatment classes in which a split has changed a response is an estimable quantity for every split in every edition, which means the fate of a distinction at the next revision may be predictable from that count and not from the difference on which the split was made.

IV.The Question That Becomes Possible

neither discipline would ask this alone

The clinical hall asks whether two subtypes are different enough to separate, and answers with a measure of difference. The phonological hall has no patients, but it settled long ago that difference is the wrong quantity, that distinctness is decided by contrast within an environment, and it has since shown on the histories of real languages that the load a contrast carries predicts whether it lasts. Neither hall asks whether the durability of a disease subtype across editions of a classification is governed by the same rule — by the count of treatment environments in which the split has changed a result — nor whether the splits that classifications cannot check, because their members never meet, behave as the rule predicts an untested opposition should.

the contrastive-load hypothesis

Does the survival of a subtype distinction across successive editions of a disease classification track its contrastive load — the number of distinct treatment classes under which exchanging the subtypes has been shown to change the response — rather than the magnitude of the molecular or clinical difference between the subtypes, and are distinctions whose subtypes stand in complementary distribution, never having shared a treatment environment, revised at the next edition without regard to either?

why it would matter

Because indication choice is made against the current edition of a classification, and a programme that enrols by a split with no load is paying for a boundary the next edition may erase, while a programme that lumps across a split with high load is enrolling people its drug cannot help. If the hypothesis holds, the load of every split could be tabulated in advance, a classification could mark which of its oppositions are findings and which are defaults, and a sponsor could choose, for a drug of a given class, the archiphenotype that class actually sees rather than the finest split the pathology permits. It would also give the lumping-and-splitting argument a quantity to argue about that is not the size of a difference.

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 trial. It needs the successive editions of the classifications that have been revised most often — the WHO classifications of haematolymphoid and of central nervous system tumours, the lung and breast classifications, the reissued consensus criteria of rheumatology and respiratory medicine — together with the labels and the randomised trials that record which treatments worked differently on which side of each line.

the honest status

Note the narrowness. The claim concerns subtype distinctions inside classifications that have been revised at least twice and for which a treatment record exists, and it concerns one outcome — whether a split survives the next edition — chosen because it is on the public record, not because it is the thing a programme finally cares about. It does not claim that molecular difference is uninformative; it claims that difference is the wrong quantity for the question of whether a split is an indication, as it is the wrong quantity for whether a sound is a phoneme. It does not claim that untested splits are false; it claims that their status is undecidable and should be marked so, which is a smaller thing. And it says something a classification committee may not welcome: that a number of the entities in its current edition are archiphenotypes waiting for a drug that can see the difference, and that until one arrives the finest split is the least useful one. It is worth saying what does not transfer. A language's environments are given by the language; a disease's environments are made by the drugs that exist, so the set that decides distinctness is itself the output of the industry that consults the classification, and a rule that was descriptive in one hall is partly self-fulfilling in the other. The whole thing rests on a comparison the editions already permit: whether load, not difference, predicts which splits are kept. If it does not, this is a well-dressed restatement of the known fact that useful subtypes are useful, and should be said so. Until that comparison is made, this remains exactly what its stamp says: a candidate. A structural echo between two halls of the library, waiting for someone with the tools of both to walk the corridor.