ADR-045: A claim about the signal is admissible however it was measured, including on people

Context

LIT-410 is a paper in Language about English syntax. Four preregistered surveys, 755 participants, a billion-word corpus study, and not one sentence about a machine. Filing it on request turned up a boundary nobody here has drawn, and the first attempt at drawing it was wrong in a way worth recording.

“Outside computer science” was never the boundary, and the corpus says so. The record holds Physical Review Letters, Reviews of Modern Physics, four volumes of IEEE Transactions on Automatic Control, SIAM, and a 1963 Proceedings of the AMS; 27 notes carry no arXiv identifier at all. Nobody argued about those, because each is machinery ML work runs on — and each is cited by work the record files, which is ADR-032’s stated trigger. Nothing here cites Goldberg and Shirtz, so on that trigger’s letter this paper is out.

The first attempt therefore reached for an exception: admit it as a LIT note that sources nothing, on a narrow rule about measuring what a topic’s practices presuppose. That framing survived about as long as it took someone to say the paper is a theory paper.

Which it is, and the record’s own test says so. CLAUDE.md: if it tells the reader what to do it is a practice, and if it says what is true it is a theory. The test is about the kind of claim and mentions no apparatus. The reflex that made this look hard — the subjects are people, so it cannot be about our systems — is a claim about where the measurement happened, not about what was claimed. And what was claimed is that a lexical unit can be an arbitrary-length phrase, delimited by construction rather than by frequency. That is a fact about language, and language is the signal, which is the thing every practice under representation-and-encoding acts on.

ADR-026 already draws this distinction on the other axis: a preconditioning scheme discovered in diffusion takes its kind, not its domain. Domain of discovery does not determine filing there, and apparatus of measurement should not here.

Decision

A document is filed by the kind of claim it makes, not by the instrument that produced it. A finding about the signal — what the data is like — is a THEORY and sources a LIT note, whether it was measured on a GPU, in a corpus, or on 755 people.

Two conditions, and the second is the load-bearing one:

1. The claim is about something the record’s practices act on: the signal, the systems, or the training process. Not about the measuring population. “Where a lexical unit ends” qualifies; “how children acquire syntax” does not, because no practice here acts on it.
2. The document states what it does not license, explicitly — and for work like this, that has to include this was not measured on a model. The failure mode is not the filing. It is a later pass reading a survey result as a result about transformers, which is the LIT-052 and LIT-025 failure with a new surface: two notes whose plausible-looking takeaways turned out not to be about their papers, and which eight practices were built on.

Such a claim takes signal-structure as its topic (v3). The word is for what the data itself is like, independent of any model: the statistics and structure of language, images and other signals, which methods exploit or have to respect. That covers frequency spectra, heavy-tailed and Zipfian distributions, and what counts as a unit. v1 filed THEORY-022 under representation-and-encoding and said in its own cost paragraph that this was the closest word available, not a good fit. Under ADR-059 that sentence was the finding, and the topic is the answer to it.

What this does not do is lower the bar for a practice. A finding about the signal underwrites a recommendation only when someone has measured the cost of acting on it, which is a separate document and usually a missing one.

Alternatives

• Decline it, and say so once. The cheapest reading, and the one DP-009 names as self-confirming: declining leaves the vocabulary looking complete because the evidence that would have shown otherwise was turned away.
• The exception framing — admit it as a LIT note that sources nothing. What this decision replaces, and it is worth recording because it looked right. It treats the apparatus as the thing in question, so it has to build a special rule to get around it; the ordinary rules already work once the claim is classified correctly. A scheme that needs an exception for a well-evidenced finding about its own subject matter has mis-stated its subject matter.
• File it and source a practice from it. The way to make the filing self-justifying, and a fabrication: the paper ran no model and recommends nothing, so a practice would be this record’s inference wearing the paper’s citation. ADR-017 is the rule against exactly that.
• Add a topic for it. Rejected in v1 and v2, and taken in v3. The argument was DP-008’s corollary: a category added to admit a single document is how a scope stops being one. ADR-059 answered that the topics are not a scope, so there was no scope to stop being one. The word chosen is broader than the v1 candidate, linguistics. It names the kind of claim (a property of the signal), not the field that measured it, which is this decision’s own rule applied to the vocabulary.

Consequences

What it buys. The boundary is a sentence somebody wrote rather than a property of what the corpus happens to contain — the DP-008 failure, avoided from the direction ADR-032 avoided it from. And it is the general rule rather than a carve-out, so the next such paper is an ordinary filing decision instead of a second exception.

What it costs. In v1, representation-and-encoding held a theory about the structure being encoded rather than about encoding it for a network, with the tag recorded as the closest available, not a good fit. v3 fixes that instead of recording it (see the history). The theory keeps representation-and-encoding as a second topic, because the practices under that word act on the structure this claim describes.

What it obliges. Counting, and a specific piece of work. The bridge from a claim about the signal to anything a model does is empirical, and for this claim the record does not hold it — the detokenization and inner-lexicon line (ARXIV-2410.05864, ARXIV-2406.20086) is unfiled. A record that admits signal claims and never files the work connecting them to models ends up with a shelf of true and inert facts.

Why it is Proposed. One document is not a pattern. What would settle it is a second finding about the signal, measured outside the ML literature and filed under these conditions — or a case where condition 1 does real work by excluding something. If neither arrives, this describes a single exception after all and should say so.