ADR-058: A library is evidence about consensus, and its observation lives in a report

Context

#183 wants the record to learn from what the big repositories implement. #287 piloted one — vLLM, twenty-one features against 351 practices — and the question it could not answer from the armchair was where the output goes.

Three candidate homes existed. The pilot and a survey of the corpus eliminated two of them on facts rather than on taste.

implementations: is not available. It is declared in no scheme, carries data on 206 of 351 practices, and holds 85 distinct values across 84 documents — essentially no reuse. Those values are four different kinds of thing: models (llama2, PaLM, DeepSeek-V3), methods (NeRF, CycleGAN, Progressive Distillation — which are not implementations at all and have their own notes), libraries (vLLM, diffusers, bitsandbytes), and prose that escaped into a list field (“the 4-bit path in the major PEFT libraries”). A field meaning four things cannot be read by anything, and a fifth meaning would settle that permanently. It is #293.

A library is not a LIT note. A note names a source that resolves and makes a claim that stays put. A repository has no fixed claim and changes weekly. Several of these repositories do have papers — vLLM’s is LIT-112, Megatron’s is LIT-043 — and those are already notes. The paper and the library are different objects, and only one of them is a publication.

Decision

A repository that implements a technique is evidence about consensus, and about nothing else. Not status, which is the record’s own judgement of whether to recommend the thing, and not the practice’s correctness. DP-005 already says adoption is not evidence; what makes consensus the exception is that its vocabulary is about adoption by construction — “converged: the field agrees and dissent is marginal, whether or not each adopter made the choice deliberately”, “universal: not doing it is what needs justifying”.

The judgement goes in consensus_note, and names what was checked. Which repositories, and when. A note saying “widely adopted” is not auditable; one saying “implemented in vLLM, SGLang and TensorRT-LLM as of 2026-09” can be disagreed with by someone who looks.

consensus_note should be declared, and cannot be yet. It is used on 206 practices and appears in no scheme — the same defect as implementations: — and this decision is what makes it load-bearing. Declaring it was attempted here and refused: luria types a field by the constraint it carries, and an optional prose field carries none, so blurb is unreachable without attaching required, many, unique or required_when — every one of which is false of this field (LU-#319).

The honest declaration is required_when on every non-default consensus: a practice asserting the field has converged should have to say on what grounds. That is 8 documents away from being satisfiable — 183 of the 191 practices with a non-default reading already carry a note — and those eight are asserting universal, converged or emerging on nothing at all (#294). Writing eight consensus assessments is editorial judgement about what the field believes, which is the highest-stakes claim this record makes, and it does not belong as a side effect of a schema change.

The observation goes in a generated report (#291), which fetches at a pinned ref, not main. This is the split the record already runs on: a source holds what a person concluded, a view holds what a machine derived (ADR-004). A citation to a moving target silently becomes false, and the pilot made that concrete — it fetched vLLM’s feature matrix from main and cannot now say which main.

Removal is a prompt, not a verdict. The report’s gone since last run column is the one signal nothing else in this project can produce, and it means “read this again”, never “lower the consensus”. A technique can leave a library because the library narrowed its scope.

Alternatives considered

• Extend implementations:. The obvious move and the one the survey killed. Adding libraries to a field that already holds models, methods and prose fragments does not give libraries a home; it gives the field a fifth meaning and removes the last chance of it ever being checkable (#293).
• A new libraries: field. Clean, and premature. It would be the third undeclared-or-barely-declared list field on this scheme, and the pilot did not show that the list is what a reader wants — it showed that a count is. A count belongs in a report, and a field would duplicate it in a place that goes stale between sweeps.
• The library becomes a LIT note. ADR-015 put infrastructure in the record, which makes this tempting. It fails on source: a note must name something that resolves and holds still, and README.md on main does neither. The paper, where one exists, is the note.
• consensus_note prose alone, with no report. What the record does today. The pilot’s output was 21 features, 14 matched, 7 absent — a number that cannot be recomputed, diffed or re-run if it lives in sentences spread across 206 documents. DP-009 is that counting is the antidote, and prose is where counts go to die.
• Status quo. The sweeps in #289 and #290 would produce findings with nowhere to land, and the first one to be written down would set the convention by accident.

Consequences

#289 and #290 are unblocked, and #291 now has a specification: pinned refs, a diff, and a gone column.

consensus_note becomes checkable and gains a burden. Naming the repositories and the date is more work than “widely adopted”, and it is the whole value — the pilot declined to move SOTA-105 and SOTA-113 precisely because one repository could not honestly support the sentence.

This decision does not fix implementations:, and routing around a broken field leaves it broken. #293 holds that, and this ADR is evidence for it: the field’s condition was discovered by needing it and finding it unusable.

promote_when is the next one to look at — declared, but only as a required_when trigger with no vocabulary, on 128 practices. Not this decision’s business; noted because the same survey found it.

Three findings came out of trying to write this decision down, which is the argument for writing decisions down: implementations: means four things (#293), eight practices assert consensus on no grounds (#294), and luria cannot declare a field it is not also checking (LU-#319). None was visible from the pilot; all three were found by asking where one sentence should live.

A report of what libraries implement will read as a leaderboard, and that is the risk worth naming. It is a record of what is served, which is not a record of what is good, and the report’s own prose has to say so where a reader will hit it — the way unbound-lineage.md says what an unbound row usually means.