Context
ADR-026 made the two topic lists identical and said why: LIT and SOTA
“differ in what a document is rather than in what it may be about.” It
added generative-modeling and vision-and-graphics to the practice
registry and recorded, in its own Consequences, that “the two vocabularies
become identical.”
They did not become identical. The keys did. Each scheme kept its own copy of the glosses in its own file, and ten of the thirteen blurbs differ — not by a word, but substantively:
| topic | practice registry | reading list |
|---|---|---|
inference-optimization | batching, cache layout, quantization procedure, sampling algorithms | serving, quantization, compression, sparsity, distillation |
analysis-and-evaluation | what to measure, what a measurement cannot tell you, which comparisons are unsound | theory, interpretability, benchmarks, measurement, debugging |
Some of that is drift nobody chose. The reading list’s own header still described a vocabulary of “the seven a practice can carry, plus five the corpus needs and the practice registry never did” — twelve words, when both lists have held the same thirteen since ADR-026. That sentence was false and had been for as long as nobody compared the two files, which is what two files makes easy.
Some of it was deliberate. ADR-027 added analysis-and-evaluation to the
practice registry and narrowed its blurb on purpose: “There, it covers a
paper’s subject and reasonably includes theory and interpretability. Here it
should cover an action.”
The move to a single luria.yaml (LU-ADR-098) made the
question unavoidable. A vocabulary is now one table, named by both schemes,
holding one label and one blurb per word. It cannot hold two glosses.
Decision
One topics vocabulary, glosses included. Both schemes name it. The
blurbs are the union of what the two files held, and they describe the
topic rather than the kind of document filed under it.
ADR-027’s narrowing does not survive, and that is the part of this worth arguing. Three reasons:
A blurb is a filing aid. It answers “which of these thirteen words is this about?”, and ADR-026 already settled that the answer does not depend on which scheme is asking — that is exactly what “differ in what a document IS, not in what it may be ABOUT” means. A gloss that varies by scheme makes the shared axis two axes wearing one name.
What a document is is already stated once, by the scheme. SOTA documents
are recommendations; LIT documents are papers. Restating that inside
thirteen blurbs copies the scheme’s own definition into thirteen places, and
copies drift — which is not a prediction here but a measurement: ten of
thirteen, plus one header that had gone false.
The narrowing was not what did the work ADR-027 wanted. Its purpose was to make four measurement practices fileable, and the topic’s existence did that. Its narrower gloss was meant to stop practices being filed there for theory or interpretability reasons; ADR-026’s filing rule already does that work generally, for all thirteen words — take a domain topic when the claim is about the domain as such — and a document that is only about interpretability is not a recommendation, so it is not a practice at all.
analysis-and-evaluation therefore reads:
how to find out whether something worked — what to measure, what a measurement cannot tell you, and which comparisons are unsound; theory, interpretability and debugging belong here too
ADR-027 is Superseded by this decision. Its substantive holding is
carried forward unchanged: analysis-and-evaluation is a practice topic,
it is a kind of claim rather than a domain, and the four practices it was
written for stay filed under it. Only the scheme-specific gloss goes.
Alternatives considered
- Keep two vocabularies with identical keys. Honours both ADRs exactly and needs no decision at all. It also keeps the structure that produced the drift, and re-adopts it deliberately at the moment the config finally offered a way out. The false header is the argument: nothing caught it for as long as it was true that two files held the same words.
- Share the keys, write scheme-neutral blurbs, and move the practice-side
narrowing into the
SOTAscheme’s own configuration comment. This was the alternative most nearly taken. It preserves ADR-027’s intent and puts it where a filer of practices reads it — and the comment now survives the migration, so there is somewhere for it to live. Rejected because it keeps the claim that the same word means two things while removing the only place a reader would compare the two meanings: worse than either honest option. If the distinction turns out to matter, the filing rule is where it belongs, stated once for all thirteen words. - Take the practice gloss for both. The narrower one is the better writing, and three of the practice blurbs lead with a definition rather than a keyword list. But the reading list genuinely holds papers about theory and interpretability — roughly a third of the corpus produced no recommendation at all — and a gloss that excluded them would misfile real documents. Range is the direction ADR-026 chose, twice.
Consequences
The drift cannot recur. There is one table, so a divergence would be an edit to a single line rather than two files agreeing to disagree quietly.
One false sentence is deleted and three stale ones corrected. The reading
list’s “seven plus five” header goes. The practice group’s comment said
“Seven categories” when there have been thirteen since ADR-026. Two
configuration comments still named the temporary codes ADR-027 and ADR-026
carried before they were numbered, which luria link --fix does not reach
inside a configuration file.
Published view paths move, as a consequence of the migration rather than
of this decision: per-status pages render under docs/<scheme>/status/
instead of docs/<scheme>/statuses/, because the directory is named for the
field and the field is status. Four directories. No link inside this record
points at the old paths — every one of them is generated — so nothing here
breaks; inbound links from outside will. The stale directories are removed in
the same contribution.
What is now harder: a scheme that genuinely needs its own gloss for a shared word has nowhere to put it, and would have to either take a different word or argue this decision back. That is the intended cost. The cheap version of that argument — quietly editing one of two files — is the thing being removed.