ADR-033: Readings made in another context port into the record; their agenda does not

Context

Seventy-four paper notes arrived from a separate research project — one about decentralized and communication-efficient training, gossip SGD, Langevin dynamics and the mean-field theory of wide networks. They are YAML, in that project’s own v2 schema, and each carries contribution, key_insight, assumptions, key_results, claims, method, concepts, connections, recommendations, research_implications, limitations and open_questions.

They are not a secondary source. They are this record’s own readings, made in a different context — the same work that produces a NOTE here, done for a different project, against a different agenda.

That is legible in the schema itself. Set the incoming fields beside the section headings this record’s NOTE documents have used since ADR-025 and they are the same list, in the same order:

incoming field	NOTE section
contribution	## Contribution
key_insight	## Key insight
assumptions	## Assumptions
key_results	## Key results
claims	## Claims
method	## Method
concepts	## Concepts
connections	## Connections
recommendations	## Recommendations
research_implications	## Bearing on the record
limitations	## Limitations
open_questions	## Open questions

Eleven of the twelve are the same field under two names. The twelfth is the one that differs in substance rather than in spelling, and it is where the whole decision lives.

Decision

Port the readings. Retarget the one field that answers someone else’s question. Take every identifier from a resolver.

1. Seventy NOTE documents, Read. Eleven sections transcribe. The order is the schema’s own, so nothing is rearranged and nothing is summarized away — theorem statements keep their rate expressions, claims keep their strength and support, assumptions keep the ones that do not hold in a colocated cluster.

2. Sixty-seven new LIT notes, Active but for one. The papers the record did not hold. Six were already here; four of those six had no reading and now have one.

3. research_implications is rewritten, not transcribed, into ## Bearing on the record. The source field answers five questions from that project’s synthesis/research_agenda.md — critical batch size, fault tolerance, cluster topology, Langevin versus large-batch, power smoothing. Those are good answers to questions this record did not ask. The section that replaces them answers the question this record does ask: what may a practice here rest on, and does one exist yet. Seventy of those were written fresh.

4. Twenty-three of the 187 recommendations are marked as not filed, in place, with the reason. Seventeen name that project’s own system and experiments by name — “when fitting MSE curves from our gossip experiments”, “map the GELS gossip coupling strength to the Kuramoto coupling K” — despite the schema calling the field project-agnostic. Six are advice to a theorist about proof technique: use the bounded-Lipschitz metric, check a log-Sobolev inequality uniformly in N, work in a rigged Hilbert space. They stay visible in the note, struck rather than deleted, because a reader who disagrees with the cut should be able to see what was cut.

5. Every title, author, date and identifier comes from arXiv or CrossRef. Not from the reading. This is the part of the earlier, more suspicious decision that survives contact with the truth, and the reason is four findings from checking all seventy-four against the resolvers:

• 1805.01361 carries a reading of Mei, Montanari and Nguyen. That identifier is a paper on hyperspectral water regression. The reading is real and is of 1804.06561, which arrived in the same batch under its own reading — so the two are merged into one note, and the wrong identifier is not filed.
• 1906.08632 carries a reading labelled with the title of Goldt et al.’s hidden-manifold paper. The reading is of the paper the identifier names — their teacher-student dynamics paper — so the identifier is right and the title was wrong. Filed under arXiv’s title, with the discrepancy in the note.
• 2102.11742 carries a reading that binds to no paper: its title and one of its authors belong to none, and its own first paragraph says the work it means is three others. The paper is filed at Deferred with no note. This is the one case where absence is the right answer, and ADR-025 already makes absence mean unread.
• Wolfowitz (1963) is given a journal, a volume, a page range and a DOI. The DOI resolves to nothing; the journal is wrong. That is a citation invented rather than mistaken, and it is the one finding here that bears on trust rather than on labelling.

Three identifier errors in sixty, and one fabricated citation in fourteen. The content is this record’s own work and ports; the bibliography around it does not, and is replaced.

Consequences

The record gains a branch of the literature it did not have, read. Local SGD and its descendants, decentralized and gossip SGD, Byzantine-robust aggregation, gradient compression, volunteer-scale training, the mean-field account of wide networks, the teacher-student line, permutation symmetry, and the consensus results underneath all of it — 67 papers, 70 readings.

## Bearing on the record is now a section that can be wrong, and that is the point. Transcribing the source field would have produced seventy paragraphs about another project’s experiments, all of them true and none of them about this record. Writing seventy fresh ones produces seventy claims about what this record could assert and does not, each of which a reader can check against the registry.

The earlier decision in this contribution was wrong and is replaced. It read the batch as a secondary source of unknown provenance, filed all 68 papers at Deferred, filed no readings, and sent all 187 recommendations to a backlog issue on the grounds that a practice may not be sourced to an unread note. The rule is right; the premise was not, because the papers were not unread. The three identifier errors that decision leaned on are real, and they are labelling errors in this record’s own filing rather than evidence about a stranger’s care — which is a reason to rebind them, not a reason to distrust seventy readings.

Two papers already had readings and did not get a second. LIT-017 (McCandlish et al.) and LIT-028 (Kaplan et al.) carry NOTE-062 and NOTE-017. The incoming readings of both are largely about fitting curves from another project’s experiments, so nothing was lost; but the record now has two readings of two papers in two places, and only one of each is filed.

Alternatives considered

File the papers unread, as the first pass in this contribution did. It asserts something false — that nobody here has read them — in order to avoid asserting something uncertain. Deferred is for papers nobody got to, and these are not those.

Transcribe research_implications as-is. It is the cheapest option and it imports another project’s agenda into this one’s documents in the voice of findings. A reader would have had no way to tell which of five questions a paragraph was answering, or that the questions were not this record’s.

Port the readings but not the recommendations. Splitting them would have put the advice a paper gives somewhere other than the reading of that paper, which is exactly the separation ADR-025 created the NOTE scheme to avoid.

Drop the 23 excluded recommendations silently. Struck-in-place costs three words per line and leaves the cut auditable. Deleting them makes the note look like the paper gave less advice than it did.