The Copy That Counts for Nothing

I.The First Known Thing — the witness that is struck out unread

textual criticism · what survives of a text is a family of copies

No ancient text survives in its author's hand. What survives is a scatter of copies of copies, disagreeing with one another in thousands of places, and the editor's task is to decide what stood in the lost original. The nineteenth century turned this from taste into method. Copying introduces errors; errors are inherited; so the shared errors of a group of manuscripts betray a shared ancestor, and from the pattern of agreement in error a family tree of the surviving witnesses can be reconstructed. Paul Maas set the procedure out formally in 1927, and the tree it produces — the stemma codicum — is not an ornament. It is the instrument that converts a heap of documents into an argument about a text.

the elimination rule, and what makes it strange

The first operation the method performs on the heap is a deletion, and it is the operation worth carrying out of this hall. It is called the eliminatio codicum descriptorum: the elimination of manuscripts shown to be copied from another manuscript that also survives. Such a copy is removed from the apparatus entirely. Not downweighted — removed, and removed before its readings are examined. The reasoning is that everything in it is either inherited from its exemplar, in which case consulting the exemplar is better, or introduced by its own scribe, in which case it is not evidence about the original at all. A perfectly executed copy and a careless one are eliminated on exactly the same grounds, because the grounds have nothing to do with the quality of the copy. They have to do with where it stands in the tree. This is why the discipline's standing maxim is that witnesses are weighed and not counted: ten agreeing manuscripts descended from one surviving exemplar are, as evidence, that one exemplar, and the reader who is impressed by the ten has been impressed by the copying rate of a scriptorium.

and what the discipline concedes it cannot do

Two further pieces of self-knowledge belong with it. The first is contamination: a scribe who works from more than one exemplar, correcting one against another, produces a witness with several parents, and a tree is then no longer the right shape for the evidence. Maas's judgement on this was that no remedy for it had been found, and it is still the standing difficulty of the field. The second is Joseph Bédier's objection of 1928, that published stemmata are suspiciously often two-branched — far more often than any plausible history of copying would produce — because a two-branched tree leaves the editor to arbitrate between the branches, which is to say it leaves the editor free. The genealogy, Bédier's point runs, is built by somebody who already has a preference about the readings. The methodological consequence the field drew is that the tree must be constructed from the pattern of errors alone, without reference to which readings the editor would like to win.

II.The Second Known Thing — the target that everyone confirms

drug discovery · target selection and target validation

Walk to a hall with no manuscripts in it. Before a programme is started, a protein is nominated as the thing to be drugged, and the case for it is assembled from the literature: the reports that knocking it down reduces the phenotype, that its expression tracks the disease, that a tool compound against it does something in a model. The assembly is real work and the field takes it seriously, because the dominant cause of clinical failure is not chemistry but the target being wrong. Amgen's group reported in 2012 that it could confirm the findings of six of fifty-three landmark oncology papers; Bayer's reported that in-house work matched the published account in roughly a quarter of sixty-seven projects; the Reproducibility Project in cancer biology spent years discovering that most of the experiments it had planned to repeat could not be reconstructed from their published protocols at all, and that the effects it did manage to repeat were, on the whole, far smaller than the originals.

what the evidence is actually made of

Underneath each of those reports is material. A cell line, an antibody clone, a plasmid, a mouse strain, a patient cohort, a small-molecule probe. These are not abstractions in the way that a citation is an abstraction: they are physical things that were passaged, aliquoted, shipped between laboratories and used again. A cell line distributed from one stock to two hundred laboratories is one stock. A monoclonal antibody whose specificity was never established is a single unresolved question, however many papers have stained with it; the field's own reviews have been saying for a decade that a large fraction of commercial antibodies do not recognise what their label says, and Cellosaurus now catalogues several hundred cell lines known to be misidentified or cross-contaminated, many of which continued in use for decades after the fact was established. The apparatus for tracking this exists — identifiers for reagents, catalogue and lot numbers, repository accessions — and it exists precisely because the material has a lineage.

what is therefore not asked

What the field does with a corpus, nonetheless, is read it. The supporting papers are assessed on their quality, their consistency and their number, and the phrase that closes the assessment is that the target is well supported, with multiple independent groups reporting concordant findings. Independent, there, means a different author list. It does not mean a different reagent. Nobody constructs the tree. There is no routine step in which a supporting paper is removed from the case on the ground that its material descends from the material of another supporting paper — removed before the reading of what it found, the way a philologist removes a copy. The nearest existing practice, orthogonal validation, is a good instinct in the same direction: confirm with a second modality, a genetic knockdown against a chemical probe. But it is applied as a qualitative check on the conclusion, not as an elimination performed on the corpus, and it says nothing about the twenty papers that share one vial.

III.The Bridge

no target dossier cites a stemma

The halls share nothing. Classical philology and preclinical biology have no common literature, no shared journals and no reviews that span them; where the two have met at all it is in the borrowing of phylogenetic software to draw manuscript trees, which runs the other way and concerns the drawing, not the evidence. But both disciplines face a heap of documents that agree with one another, and both need to know what that agreement is worth. One of them decided, and the decision was counter-intuitive enough that it had to be made explicit and defended.

X: descent is assessed before content, and agreement is the cheapest thing a corpus makes

Call it witness elimination. Where a body of evidence has a descent structure — where a document's material was obtained from the material of another document in the same body — the evidential weight of the corpus is fixed by the number of its surviving independent sources and by nothing about its size. A witness that can be shown to derive from a witness still in front of you contributes nothing whatever it reports, and the demonstration is made from the lineage and not from the reading, which is what makes the operation a deletion rather than a discount. From which follows the half that biology has not ported, because only the editors have had to build a case on documents that reproduce themselves: agreement among descendants is the cheapest output a corpus can produce, so apparent convergence accumulates fastest exactly where independence is least, and a corpus therefore looks most convincing at the moment it has stopped being evidence. A widely distributed cell line is a scriptorium. A monoclonal clone used in two hundred papers is an exemplar with two hundred descripti, and the confidence a reader takes from their concordance is confidence in the shipping records of a repository. Counting the papers is not a rough approximation to weighing the evidence. It is the assumption that the papers have no tree.

IV.The Question That Becomes Possible

neither discipline would ask this alone

Biology already has the lineage data. Reagent identifiers, cell line accessions, antibody catalogue and lot numbers, plasmid repository numbers and mouse strain designations sit in the methods sections, and the registries that resolve them are public. They were built to help a reader repeat an experiment. They have not been used to build a tree over a corpus and then delete from it.

the elimination hypothesis

Can the corpus supporting a candidate target be resolved into a stemma of its material witnesses — cell line, antibody clone, construct, model organism, cohort — and does the number of witnesses surviving elimination explain which target findings withstand an independent replication attempt, where the number of supporting publications does not?

why it would matter

Because the count that is currently persuasive is the one the mechanism predicts should be least informative, and the two come apart in a direction that can be checked. A target with forty papers descending from three stocks and a target with six papers from six stocks are not close cases under the elimination rule: the second is the stronger corpus, and under present practice it reads as the weaker one. The rule also makes a claim about where surprises come from — that the failures which look inexplicable, where a well-supported target collapses on contact with a new laboratory, should concentrate in corpora whose surviving witness count is small and whose publication count is large. And it supplies a second, sharper weighting from the same hall. The philologist's lectio difficilior holds that the reading harder to explain as an accident of copying is the stronger; the biological analogue is that a witness whose modality could not have produced the finding by the same artefact as the others — a human genetic association standing beside a chemical probe, rather than a fourth knockdown in the same line — is worth more than concordance among witnesses that share a failure mode. That is a quantity that can be computed from a methods section. It is not currently computed from anything.

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 is retrospective first, and the data for it has already been collected by people who were asking a different question.

the honest status

Note the narrowness. The claim is about corpora of preclinical findings with recoverable material provenance, and about the number that should be read off them. It says nothing about whether any particular target is real; a corpus with one surviving witness may be reporting something perfectly true, and the index is a property of the evidence, not of the biology. It makes no prediction about clinical outcome, which has too many other causes between it and here. And there is a disanalogy at the centre of it that could prove fatal rather than merely limiting: a scribe's copy adds no information about the original, whereas a second laboratory using the same cell line has genuinely run a new experiment — the material is shared, the hands and the conditions are not. Whether that intermediate status is a difference of degree, which the weighting can absorb, or a difference of kind, which would void the elimination, is not something this candidate settles. Until a tree has been built over a real corpus and counted against its replication record, 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.