ADR-017: What source: holds, and where the work it excludes goes

Context

ADR-010 made source: a list, and its argument was countability: a reader should be able to tell “rests on one unreplicated claim” from “rests on four independent groups”. It did not say what fills the list.

Two conventions grew in the gap, and both are defensible readings.

SOTA-132 counts adopters as support. Its source: comment says so:

Primary: the controlled comparison against full attention. The rest is what corroborates it — the shipped 3:1 layouts and the module they use (ADR-010).

SOTA-150 counts adopters as consensus data. Its source: holds five origin papers; the adopters are in its consensus_note:

Four labs and every frontier model the record holds from 2024 on: DeepSeek (LIT-160, LIT-139), Moonshot (LIT-132, LIT-131), Alibaba (LIT-182, LIT-136) and NVIDIA (LIT-183).

Under two conventions the list is not countable, which is the property ADR-010 was filed to get. SOTA-150 reading converged on five sources and SOTA-132 reading its value on six mean different things, and only a reader who opens both can tell.

The cost came due in #58. That pass needed a rule and there was none, so one was invented mid-pass — if this paper were retracted tomorrow, would the practice need rewriting? It works. It was also nowhere a contributor would look, and thirteen practices were corrected under a rule that existed in a commit message.

Two attempts to mechanise it failed, and the failures are the useful part. Measured against the corpus:

candidate check	fires	why it is wrong
codes in consensus_note / promote_when ⊆ structured fields	22	SOTA-122’s promote_when names LIT-153 precisely to say it does not count
codes in the ## Source section ⊆ source:	17	SOTA-124’s Source section names LIT-166 and LIT-175 as the positions it argues with

Neither violation is always wrong. A check written before this decision would have enforced whichever convention its author held and generated false findings against the other half of the corpus.

Decision

source: holds the work that produced evidence about the claim. The test, promoted from #58’s practice to the record’s rule: if this document were retracted, would the practice need rewriting? That admits the origin of the claim and independent corroboration of it. It excludes contrast, alternatives, components that have practices of their own, and asides.

Adoption without a test is consensus data, not support. This is the disagreement above, and consensus.yaml already contains the answer. Its blurb for converged reads:

the field agrees and dissent is marginal, whether or not each adopter made the choice deliberately

A non-deliberate adopter is evidence about the field, not evidence about the claim. Ten laboratories shipping a design is not ten results; it is one belief, held ten times. So the line is not who published but what they did: a report that ran the comparison is a source, and a report that shipped the design is consensus data. Kimi Linear ablating NoPE against RoPE at matched configuration is a source. Qwen shipping a 3:1 interleave without reporting the alternative is not.

consensus_note may name work absent from source:, and that is the division of labour rather than a defect. It follows that ADR-010’s countability claim needs restating: consensus is countable against sources and adopters, not against source: alone. What makes it countable at all is a group count, which this record does not store — filed upstream as luria#209.

The ## Source section is argument, not enumeration. It says what each source contributes and how they fit, and it may name non-sources provided it names them as such — SOTA-124 naming the position it argues with is correct prose and correct filing at once. So the prose is not a subset of the field.

The field is a subset of the prose. Every declared source must be discussed somewhere in the body. A source the document never explains is a source no reader can evaluate and no later contributor can re-triage, and it is the one direction that is always wrong.

Where the excluded work goes

Defining source: narrowly evicts something, and the eviction is the half of this decision that was nearly left undecided. The literature that set up the problem is not evidence for the recommendation, so it leaves the field — and lands in prose, which is the failure this record has spent a week correcting.

The record already holds the answer and cannot say it. Seven of the nineteen extends: edges in the literature mean “this exists because the parent is broken”, and every one of them states the defect in its summary:

edge	the defect, in the child’s own summary
LIT-140 → LIT-141	hyper-connections “break the identity mapping a residual guarantees, which destabilises deep stacks”
LIT-151 → LIT-140	mHC’s constraint “bounds it from above but not below, so the mixing can only shrink”
LIT-181 → LIT-140	”three named defects in mHC’s doubly-stochastic constraint”
LIT-192 → LIT-045	RoPE does not extrapolate, so interpolate “instead of extrapolating past it”
LIT-193 → LIT-192	interpolation “destroys the high-frequency components and stalls”
LIT-200 → LIT-030	SwiGLU’s x² “is what enlarges the output range and produces outliers”
LIT-210 → LIT-045	”the answer is not the one RoPE’s own paper gave”

The other twelve are cumulative — Mamba-2 from Mamba, GShard from the sparsely-gated layer, Switch from GShard. One relation carries both, so the lineage page renders “builds on its success” and “exists because of its failure” identically, and the difference survives only as a sentence in each child. That is ADR-011’s own observation, unnamed: it records that pointing oHC and sHC at mHC rather than at Hyper-Connections was a real decision, because “both take mHC’s design and replace its constraint”.

Succession gains a sign. A work that names a defect in its parent as its motivation is correcting it; a work that does not is building on it. The test is the child’s own text, not an inference about what its authors were thinking — which is why this is not called inspiration.

The problem-setting work that succeeds nothing is a different relation, and it is not this one. NoPE exists because explicit encodings fail at length generalization, and it builds on none of them; that is “rival answers to the same question”, filed upstream as luria#208 and deliberately not reinvented here.

How the sign is carried, and why not faceting

Faceting (LU-ADR-076) classifies a document — a frontmatter field backed by a scheme-local vocabulary, which facet_by renders beside each chain step, and which is how this record’s lines already show (Active, emerging). A sign on succession classifies an edge. Three ways to close that gap, and the choice matters more than it looks:

A document facet — succession: corrective | cumulative. Available today with no change to luria at all: a succession.yaml, one luria.toml clause, and the chain page renders it through facet_by unmodified. It is also structurally wrong, and the record says exactly when it breaks: it types the document, so a work that corrects one parent while building on another cannot say so. Today LIT-137 is the only document with two parents and it is cumulative to both, so the approximation holds — on one document’s worth of luck.

A second relation — corrects:, converse corrected_by:. Edge-level by construction, because the sign is which field the code sits in. It never engages ADR-011’s prohibition on qualifying a relation, since no reference entry gains an attribute. It gives the trunk the converse for free, and LIT-141 reading corrected_by: [LIT-140] is the branched-conflict shape this record already says it wants to read from the trunk. Its cost is one small change upstream: a chain’s spine is a single relation today (relation: str), so two succession relations would render as two broken lines rather than one signed one.

A per-entry attribute on extends:. The correct general shape and the largest change, since reference fields hold codes and nothing else. It is also the one that has to argue with ADR-011 — though the prohibition there is against a condition beside a checked reference, which nothing evaluates, and a closed two-value vocabulary is not that.

Decision: the second relation. It needed one thing upstream — Chain.relation accepting a sequence the way facet_by already does — filed as luria#211 and shipped in 0.11.0, so relation = ["extends", "corrects"] walks both as one spine. The document facet was available the whole time and is still refused: it types the document, and a shape that is wrong the first time a paper corrects one parent while extending another is not worth the week it would have saved.

The check that follows

source: ⊆ codes cited in the body. Measured before adopting it: 150 of 153 practices already satisfy it, and all three exceptions are real defects.

• SOTA-109 sources LIT-100 and its body cites only LIT-024, the design it replaced. The paper the recommendation comes from is named in the field and nowhere in the prose.
• SOTA-121 — “use Muon” — sources LIT-159, the work that introduced Muon, and never mentions it.
• SOTA-150 sources LIT-160 and never mentions it.

Three for three, on a corpus of 153, against a rule that holds 98% of the time without anyone having stated it. That is the profile a lint wants and the two rejected candidates did not have.

Alternatives considered

Adopters count as support. SOTA-132’s convention, and the one that looks right: a design running in four frontier models plainly is evidence of something. It loses because of what it is evidence of. A converged practice would accumulate every model report that ships it, source: would become a roster, and the question a reader actually brings — which paper should I read — would drown in it. consensus.yaml’s own admission that adopters may not have chosen deliberately is the record already conceding the point.

A subset lint in the prose→field direction. The obvious mechanisation, tried twice, rejected on measurement rather than taste: 22 and 17 false findings respectively. Recorded here because it will look like a good idea again — the intuition that prose naming a paper implies the field should too is exactly wrong in a record whose documents argue with the literature.

Annotated references, so the section can be generated. The reason the Source section is hand-written is that the field cannot carry the qualifier; “Hägele et al., for the comparison against cosine” is the load-bearing half. Structuring the qualifier would make drift impossible rather than detectable. Not taken here, because luria’s ADR-011 says do not add a qualifier to the relation — a condition beside a checked reference is prose in a data field, so nothing evaluates it and nothing notices when it stops holding. A descriptive note is not a condition, and the distinction may hold, but re-opening that line deserves its own argument.

inspired_by as a new relation. The shape this arrived as, and it loses twice. It duplicates edges that already exist — in all seven corrective cases extends: is declared and the gap is that it carries no sign, so a second relation beside it would give the record two ways to say one thing and a choice to make every time. And “inspired” is a claim about causation inside someone’s head, which the record cannot check: extends, compared_against and source are all verifiable against the paper, and the discipline is worth more than the expressiveness. What the record can check is whether the child names the parent’s defect as its motivation, which is what the decision above uses instead.

Splitting affirmative and detractive on a wider set of relations. Tempting once the distinction is named, because it applies to source: too — a source can be evidence for or against. It is already there and better placed: contested_by (ADR-016) is the negative half of source:, required exactly where consensus: contested asserts it. Generalising the sign further would re-derive a field the record has.

Leave it undecided. The status quo, and it has a running cost: #58 is correcting a quarter of the registry under a rule invented for the occasion, and the next pass would invent a different one.

Consequences

The rule reverses one change made the same day it was written, which is the clearest evidence it decides something. SOTA-142 gained LIT-119 under #62 on the strength of its own sentence — Falcon-H1-Tiny “is the record’s independent adoption … which is what moves this from one group’s result to a practice”. Under this decision that adoption moves the consensus and not the evidence, because Falcon-H1-Tiny used the power law rather than testing it. Corrected in the same contribution.

SOTA-132’s shipped-layout entries are the remaining case and are left to #58, which is still working through the architecture cluster.

Three practices gain a sentence about a source they had been declaring in silence.

What this obliges: a practice that adds an adopter must ask whether the adopter ran the comparison, and the answer belongs in the prose either way. That is a judgement per source, which is the cost of having the list mean one thing.