ADR-020: The recommendations are scoped by kind of claim, not by domain

Context

ADR-003 gave the SOTA scheme seven topics and the LIT scheme those seven plus five more, on the ground that “the reading list covers ground the practice registry does not.” DP-008 then read that arrangement back as a scope statement: the recommendations are scoped to language-model training by construction, so a vision paper can be a note and cannot be a recommendation — “not on judgement, not on evidence, but because there is no category for it to take.”

Three things have since gone wrong with that.

The record already contradicts it. SOTA-113 (continuous batching) and SOTA-115 (prefill/decode overlap) are Active recommendations about serving, filed under systems-optimization. So are the record’s quantization practices. “Scoped to language-model training” was never true of the practices the record actually holds; it was true of the sentence in DP-008.

The bundling was wrong in the other direction too. ADR-003 justified all five extra LIT topics with one sentence, and it holds for generative-modeling and vision-and-graphics but not for adaptation-and-tuning or inference-optimization — the record files fine-tuning, preference-training and quantization practices under the seven, several of each. LoRA sat in this corpus for a year sourcing nothing because its tag looked like a scope boundary and was not one.

And the cost is the thing the anthology exists to avoid. A scope drawn by domain makes the record blind to work whose transferable content is exactly the kind of claim it collects. EDM’s contribution is a preconditioning scheme, a noise-level sampling distribution and a loss weighting. Segment Anything’s is a model-in-the-loop annotation bootstrap. Those are a training-optimization claim and a data-pipeline claim that happen to have been discovered in image models. Declining them keeps the record ignorant of its own subject.

Decision

Scope the recommendations by the kind of claim, not by the domain the work was discovered in. A practice belongs here if it is an instruction about how to build, train, adapt or serve a model that one of the seven topics can express. Where the work came from is not the test.

Nothing else changes. The seven topics stay exactly as ADR-003 set them — they are kinds of claim, and they were never the problem. The LIT scheme keeps its twelve. No field is added.

DP-008’s principle stands unchanged; its worked example is amended, because the example described a scope this decision replaces.

Alternatives considered

• Widen primary_topic — add generative-modeling and vision-and-graphics to the practice seven. The obvious move, and it defeats the purpose. primary_topic is exactly-one and it is the browsing axis: a diffusion preconditioning practice tagged generative-modeling renders on the generative page and not beside the LLM normalization practices it is a sibling of. Widening the topic list gives cross-domain work its own page, which is the opposite of cross-pollination — a topic spent naming the domain is a topic not spent naming the claim.
• Merge the two vocabularies. ADR-003 rejected this and was right: it “puts a diffusion sampler paper under training-optimization or model-architecture and makes both pages worse.” Nothing here disturbs that. What changes is only that a note’s topic no longer implies anything about whether a practice can be drawn from it.
• Add a domain: axis in this decision. Drafted, built, and withdrawn — see below. It is the right shape eventually and the wrong thing to build now.
• Add a phase: axis — pretraining, post-training, inference. Same answer, same reason. Inference practices already have somewhere to go: SOTA-113 and SOTA-115 have been sitting comfortably under systems-optimization.

The axis this decision does not add

A first draft of this decision added a domain: field to the SOTA scheme — a many-valued vocabulary saying what a practice is known to apply to, with domain-general as the value that marks what transfers. It is recorded here rather than in a rejected branch, because the reasoning is what stops the next person rebuilding it.

Why it was withdrawn. It lets a practice assert where it applies while naming no evidence for that assertion, in a record whose first principle (DP-001) is that a recommendation without a citation is an opinion, and which holds under ADR-017 that source: carries the work that produced evidence about the claim. The draft demonstrated its own failure: eight practices were backfilled, and three claimed domain-general — transfers across every modality — on the strength of one language-model paper each. That is DP-005’s shape, a field that feels like evidence and is not.

Three further objections, any one of which would be enough on its own:

• It would be almost entirely default. One note in 218 is a cross-domain transfer result. Across 181 practices the axis would be ~99% unassessed, and a field that is overwhelmingly default is not neutral — a reader takes absence for a value. consensus: survives that trade because assessing consensus is a by-product of touching a practice; assessing domain is not, since it means reading papers in another field.
• It aims at the weakest link. Being blind to relevant work has a pipeline: never encountering a paper, refusing it on scope, filing it as a note and drawing no practice, or filing a practice nobody browsing finds. This decision fixes the second. Issue #85 is direct evidence that the third is the binding constraint — 53 inherited notes with no practice, no reason and no retirement. A domain: field addresses only the fourth.
• A category added to admit a single document is how a scope stops being one — DP-008’s own corollary, applied to the decision that amends DP-008’s example.

The decomposition, for whoever builds it. Three different facts were bundled into the word “domain”, and they belong in different places:

the fact	its nature	where it belongs
what a paper is about	descriptive, cheap, uncontentious	the LIT scheme
where a practice has evidence	derivable, once notes carry the above	generated on SOTA, not typed
whether a practice transfers	a judgement that must name a paper	its own axis

The third has a ready-made shape: ADR-016 established that a value asserting specific other work exists must name it, which is why contested_by is required exactly when consensus is contested. A transfers: axis would work the same way — shown-to-transfer requiring a transfers_shown_by: reference — and needs no new machinery, because required_when on a reference already exists.

Note also that derivation is not available today, which is the trap in the middle row: only two of the LIT scheme’s twelve topics (generative-modeling, vision-and-graphics) carry domain information at all. The other ten are kind-of-claim words. So “derive the practice’s domain from its sources” means first adding a domain axis to LIT — the top row — and the order in that table is the order to build in.

Consequences

The bar for a new practice is now a question about the claim — is this an instruction one of the seven can express? — rather than a question about the paper’s field. That is a wider gate, deliberately.

Issue #85’s remaining surface changes shape. Twenty-three inherited notes were set aside as out of scope by construction; they are ordinary unread notes now, and the ones with a transferable claim — EDM, Segment Anything, Latent Diffusion, the distillation pair — are the most interesting reading left in the corpus rather than the least.

The first thing this decision asks for is a re-reading of LIT-090, which is the one cross-domain transfer result the corpus already holds — Lion, reported across vision, vision-language, diffusion and language modelling — and which sits Rejected on an inherited retirement its own note flags as unreviewed. It is simultaneously the evidence a transfers: axis would need and a question the widened scope makes live.

The risk this accepts: the record can become a directionless list of everything. The guard is that nothing was loosened except the domain test. A claim still has to be an instruction about building, training, adapting or serving a model, still has to take exactly one of seven categories, and still has to name a paper that produced evidence about it.