ADR-057: Three practices from one paper are siblings by source, not a line of practice

Context

The practice chain’s one unbound line, surfaced by LU-#311:

SOTA-036, SOTA-037, SOTA-038, SOTA-279, SOTA-280, SOTA-281
share no `tags` across the whole line

SOTA-036 is the sole cause; drop it and the remaining five intersect on in-context-learning. Every edge binds, so unbound-relations never reported anything — SOTA-037 and SOTA-038 both share model-architecture with SOTA-036, and SOTA-038 shares it because a previous pass put it there to bind that edge (v2, 2026-09-18, whose note reads “In-context learning is a capability of the decoder-only-at-scale family SOTA-036 describes, which is what the relation between them asserts”).

That note justifies the relation, not the document. The tag test is the document’s: would someone browsing model-architecture be right to expect “ICL permits few-shot task adaptability”? They would not. So the edge was bound by a tag that should not have been added, and the line is where that showed.

Decision

Remove both relations. SOTA-037 and SOTA-038 no longer extends SOTA-036, and SOTA-036’s extended_by is empty. SOTA-038 drops model-architecture and goes to v3.

Two facts decide it, and neither is that the line does not bind.

extends means “could not stand without”, and these do. The relation’s own blurb: “the earlier practice this one builds on and could not stand without.” SOTA-036 is a specific recipe — decoder-only, a broad web corpus, a fixed context, a single next-token objective. SOTA-038 (“ICL permits few-shot task adaptability”) does not depend on any of those particulars. Change the corpus or the context length and it still stands; it depends on the class of model, not on this recipe. The same holds for SOTA-037. A relation whose test the edge fails is a relation that is wrong, which is ADR-049’s second reading.

The connection is already recorded, by source. All three name source: LIT-035 and introduced_by: LIT-035. They are three recommendations drawn from one paper — siblings by provenance. extends was saying “same paper” a second time, in a vocabulary that means something else, and the practice chain then read that second saying as a lineage of refinement.

SOTA-279 extends SOTA-038 stays: chain-of-thought exemplars genuinely could not stand without few-shot prompting, and it is a narrower case of the same recommendation.

Alternatives considered

• Add in-context-learning to SOTA-036. The one-tag fix, and forbidden by that topic’s own blurb — “getting behaviour out of a fixed model by what you put in the context.” SOTA-036 is how the model is made. This is the failure the vocabulary’s alert names, and it is tempting precisely because it is one line.
• Split extends into a second relation — enabled_by, for “a rule that only exists because the earlier one is followed.” Considered at length, and it solves a real problem: that third sense in extends’ blurb cannot preserve subject by construction, which is why an invariant over the chain asserts more than the relation promises. It is the wrong fix here because it preserves a lineage that is not one — the honest reading of these two edges is not “a different kind of succession” but “not a succession”. The general question is worth its own decision, on an edge that survives this test.
• A domain tag — next-token-prediction — under a domain sub-vocabulary (LU-#298). It would bind all six, and the diagnostic behind it is good: where a rule exists only because an earlier one is followed, the earlier one’s subject is often the presupposition the whole line inherits. It is still the wrong fix here, for the same reason: it makes the check pass by recording what the six have in common as background, when what the line claims is succession. It would also make the invariant easy to satisfy everywhere — every autoregressive-LM practice shares that presupposition — which is the vacuity the check exists to catch.
• Keep the relations and accept the standing row. LU-#311’s lint.baseline makes this cheap, and it is the wrong use of it: a baseline is for a residue a record has read and cannot yet resolve, not for one it has read and can.
• Status quo. Leaves a tag on SOTA-038 that fails the tag test, and a practice-lines.md that shows a refinement lineage from a pretraining recipe to a prompting technique, which is not what happened.

Consequences

docs/practice-lines.md loses a line and gains two singletons. SOTA-036 and SOTA-037 now stand alone, and the chain-of-thought line starts at SOTA-038. The narrative — GPT-3’s recipe, then in-context learning, then chain of thought — is real and is not a line of practice; it is readable from LIT-035’s own page, which lists all three practices sourced to it, and from their prose.

The unbound line goes to zero, and that is a consequence rather than the reason. Had the relations survived the “could not stand without” test, the right outcome would have been to leave the row standing and say so.

SOTA-037 keeps model-architecture, which is a weaker tag than it looks and is left alone deliberately: it was original tagging rather than a tag added to bind an edge, so removing it is a different judgement made on different evidence, and there is now no relation for it to prop up.

The extends blurb still licenses three senses, one of which cannot preserve subject. This decision does not fix that; it removes two edges that were not that sense either. The next line the invariant cannot bind is where that question gets answered.