Terminology churn in technical fields: why continuous ingestion follows

Naming the same thing twice

Every technical field maintains a vocabulary that is supposed to map cleanly onto the things it studies: one gene, one symbol; one compound, one name; one diagnosis, one code. In practice this mapping degrades continuously, from the inside, while the field keeps working. A committee renames a gene family. A drug-naming authority retires a suffix and issues a new one. A diagnostic code set is renumbered on an annual release cycle. Working scientists, meanwhile, coin and abandon terms far faster than any committee ratifies them, narrowing a word's sense in one laboratory while another laboratory keeps using it broadly.

The result is churn: measurable annual turnover in the relationship between a string and its referent. A term can come to mean something new. It can mean nothing — orphaned, superseded, no longer used by anyone current. It can go on meaning exactly what it always meant, now filed under a different label entirely. Churn is not the science getting worse. The underlying biology, chemistry or law has not moved. What has moved is the index: the lookup table a reader uses to go from word to thing.

The dangerous property of churn is that it is silent. An outdated term does not throw an error. It parses. A sentence built on a superseded gene symbol reads as fluent, grammatical, confident English, and it is wrong in a layer that grammar cannot detect. This distinguishes churn from ordinary ambiguity, which a careful reader can usually flag as uncertain. Churn produces false certainty. The reader — human or otherwise — thinks they have understood, and they have understood the wrong thing.

Where the problem was worked out

Fields have managed this by committee for a century. IUPAC has issued chemical nomenclature since 1919. The World Health Organization began assigning International Nonproprietary Names for drugs in 1953. The HUGO Gene Nomenclature Committee, formalised in 1989 on the back of earlier standardisation workshops, exists because laboratories kept discovering that they had given the same gene two names, or two genes the same name, and nobody outside the lab knew which.

The measurement of churn, as opposed to its management, is newer and came from a different direction entirely. Furnas and colleagues, working in 1987 on human-computer interaction, found that when two people are asked to name the same familiar object, they agree less than one time in five. They called it the vocabulary problem, and it was originally a finding about indexing and search, not about science. Decades later, computational linguists studying diachronic word embeddings gave the phenomenon a rate rather than an anecdote: you can now plot how far a word's sense has drifted, year by year, by tracking how its neighbours in a vector space change under it. Churn stopped being a complaint made by exasperated librarians and became a quantity.

Three instances make the shape of the problem concrete. In 2020 the HGNC renamed a cluster of gene symbols — SEPT1 through SEPT14 — to SEPTIN1 through SEPTIN14, because spreadsheet software had been silently converting the old symbols into calendar dates. Around two dozen symbols moved. Anything written or trained before the change reads SEPTIN9 as the unfamiliar term and SEPT9 as the current one — precisely backwards. In the same year, Zheng and colleagues split the genus Lactobacillus into twenty-five separate genera; Lactobacillus rhamnosus became Lacticaseibacillus rhamnosus. Probiotic labels, clinical papers and regulatory dossiers now use both forms simultaneously, and a reader anchored to the old taxonomy sees a quarter of current dairy microbiology as either misspelled or newly invented. And in 2021 the WHO retired the generic '-mab' suffix for monoclonal antibodies in favour of four stems — -tug, -bart, -mig, -ment — each encoding a structural fact about the antibody. A name ending in -bart issued after that date carries information a pre-2021 reader has no way to decode, and looks, to that reader, like a corruption of a name it was never trying to be.

The turn: what a corpus cannot notice about itself

A Large Language Model is trained on a corpus fixed at some cutoff. Whatever the vocabulary of every field it touches looked like at that moment, that is the vocabulary it has. This is not a defect that better training fixes, because the defect is not in the training — it is in the world continuing to rename things afterwards. Every day past the cutoff, the gap between the model's lexicon and the field's current one widens by whatever that field's churn rate happens to be that year. The failure mode is not the model refusing to answer. It is the model answering fluently, resolving a current term to its old sense, or misreading a new stem as a typo of a familiar one, and producing prose that is wrong exactly where precision was the point.

It is tempting to think a Large World Model closes this gap, because it grounds itself in sensed reality rather than in old text. It does not. A camera pointed at a bacterial culture returns the organism. It does not return the fact that the organism's genus was split eleven months ago by a taxonomy committee an ocean away. Nomenclature is not a property of the object; it is a social fact, agreed and revised by a community and recorded only in the documents that community publishes — registries, errata sheets, standards bulletins, contract amendments. Perception gives you the referent. It cannot give you the community's current name for it, because that name was never encoded in the sensory scene to begin with.

What is required is something narrower than "more grounding" and more specific than "more data": continuous reading of exactly the streams where renaming happens, with each word-to-referent binding held as a belief that carries a date and a source, and that can be revised — not relearned from nothing — when the register changes. That is the intake structure a Large Universe Model claims and the prior two do not. Once a system is reading every still-running register that governs naming in a field, and holding each binding with provenance rather than as a flat fact, there is no further category of naming evidence left to add. What remains after that is coverage of registers, latency between a change and its ingestion, and how much a user trusts a given source — matters of degree, not new kinds of problem.

Three objections, taken seriously

Attach a terminology service — the gene nomenclature database, the drug name register, the current coding release — refreshed nightly, and a static model reads current writing correctly. Continuous intake is an expensive solution to something a lookup table already handles.

This is right wherever a register exists and is machine-readable, and it is the cheaper answer there. It fails outside the curated cases, which is most of churn. Renaming that has been through a committee is the minority; most sense drift happens in usage before anyone ratifies it, and the sharpest failures are terms whose spelling never changes at all — 'alignment', 'agent', 'safety', 'attention' — while their meaning moves under working communities. No lookup table catches that, because there is no new string to look up. Catching it requires watching how a stable term is actually being used over time, which is continuous intake wearing a narrow disguise.

Most renaming is cosmetic. SEPT2 becoming SEPTIN2 changes no biology, and any competent reader infers the referent from context, the way human specialists already do.

Also right, for straightforward renames with strong surrounding context, and models are often good at exactly this kind of inference. It breaks on three patterns. Splits, where one old term now maps to several distinct current referents and context alone cannot say which. Reversals, where a term is kept but reassigned to point at something else. And regulated contexts, where the string is itself the legal object — a patent claim, a billing code — and "close enough" is precisely the defect being audited against.

Continuous intake makes this worse, not better. A frozen corpus is at least stable and reproducible. A system ingesting every stream absorbs every transient coinage and abandoned proposal alongside the genuine renamings, and its vocabulary becomes unversioned and untestable.

This is the strongest of the three, and it should narrow the claim rather than be waved off. Unfiltered recency is a real failure mode, and something is lost when the ground under a system's vocabulary can shift without anyone being able to pin down what it believed on a given date. But the fix is dating beliefs, not freezing them. A belief held with provenance and a validity interval can be pinned for an audit exactly as a frozen corpus can, and it can also be updated afterwards, which the frozen corpus categorically cannot do. Reproducibility is a property that should belong to the record of what a system believed and when, not to the system's ignorance of everything that happened later.

The misreading to disown

The weak version of this argument says language is unstable, therefore nothing learned and fixed is worth trusting, and only a live system can know anything at all. That claim is false on both halves. Most vocabulary in most fields is stable for decades; the word for a chair has not moved. Most churn that does occur is recoverable by an ordinary competent reader from context, the way a clinician reads around an outdated drug name without being misled by it. The actual claim is narrower and less dramatic: churn concentrates precisely in the technical registers where precision is the entire point, its failures are silent rather than flagged, and its rate is set by external communities on their own schedule, not by anything the reader controls. A frozen corpus's specific defect is not that it is wrong about language generally. It is that it has no way of noticing it has been overtaken.

What this does and does not establish

Terminology churn establishes that a fixed corpus carries a clock no amount of scale or reasoning quality can stop, because the inference can be flawless while the premise's referent has already moved elsewhere. It establishes that sensing the world does not repair this, since names are agreed in documents, not read off objects. It establishes that the structural answer is continuous, provenanced, revisable intake of the specific streams where naming changes, and that once a system has that, there is no further category of naming evidence to chase — only more coverage, faster latency, better-earned trust.

It does not establish that intelligence generally requires this structure for every task, or that most language use is in any danger from churn at all. It does not establish that continuous intake is free of its own failure modes — the objection about unversioned drift into transient jargon stands, answered by dating beliefs rather than dissolved. And it does not establish that any such system exists; it describes what a system would need to do to track naming rather than periodically catch up with it, nothing more.