ADR-052: Two more topics: one that had a trailing clause for a home, and one that had nothing

Context

The papers-feed worklist is ranked by revisits, and the top of it had drifted away from what the sixteen topics could express. Working down it meant repeatedly skipping the most-read entries — including the single highest, at 22 revisits — and filing the in-scope items beneath them.

That is the pattern DP-009 names and warns about. A document with a real claim and no category has two readings: the claim is out of scope, or the vocabulary is short a word. The first is cheaper, requires no work, and is self-confirming — decline the document and the vocabulary keeps looking complete, because the evidence that would have shown otherwise has just been turned away.

Asked to extend the vocabulary, the question was which words, and the instrument is DP-009’s: count, then read the hits.

Decision

Two topics, taking the vocabulary from sixteen to eighteen.

multimodal-learning

What changes when one model has to take in more than one kind of signal — joint architectures and fusion, contrastive image-text training, cross-modal transfer, and what a second modality buys the first.

This is a naming gap, and it has the exact shape ADR-050 found for numerics. model-architecture’s blurb ended with “multi-modal designs” — a trailing clause on a topic about the shape of a network, doing duty for a subject about how many kinds of signal go into it. Nine documents sat across six topics, with model-architecture holding three of them.

Thirty-five more are queued on the incoming list, at 55 revisits between them, and they are not the generative ones: text-to-image and text-to-video work already has a home in generative-modeling, whose blurb names it. What is left is vision-language encoders, multimodal foundation models, fusion architectures, image-text datasets and cross-modal behaviour.

model-architecture’s blurb loses the clause, and now says what it is about: how the network is shaped, not how many signals it takes in.

deployment-and-society

What a system does once it is running among people, and how anyone outside it could find out — audits of deployed platforms and recommenders, information ecosystems, moderation and governance, and the access problems and funding pressures of studying any of it from the outside.

This is a scope extension, and it is the only topic in the vocabulary that named nothing on the day it was added. A probe across all 749 record documents returned one hit and that hit was a homonym. The incoming list holds twenty-four, led by the highest-revisit paper on it.

The blurb’s last clause is deliberate. A quarter of those twenty-four are methodological rather than substantive — how to do this research, what happens when platform data access is withdrawn, who funds the high-profile studies, what a benchmark for societal-impact work would have to look like. Those are claims of a kind this record already files: recommendations about how to measure. ADR-026 is the rule that lets the rest in — a claim takes its kind, not its domain.

Adding a topic that names nothing is the weakest move in this decision and it is made deliberately, because the alternative is the self-confirming decline. The first filing follows this contribution; if it does not, this term is the one to reconsider.

Alternatives considered

Six candidates were counted and rejected. Recording them matters more than the two that passed, because the next pass should not re-measure them.

• Mechanistic interpretability. 19 apparent record documents, about ten real after reading — “circuit”, “feature” and “interpretable” are used loosely across the corpus, and an evolution-strategies paper was among the hits. The survivors sit in analysis-and-evaluation and representation-and-encoding, and both blurbs already name the work: the first says “theory, interpretability and debugging belong here too”. Two honest homes is not a half-home.
• The physics of learning. 28 record documents — and 17 of them already take analysis-and-evaluation as primary, 61%. By the test ADR-050 set, a subject the vocabulary can say lands mostly in one topic, and a subject it cannot scatters. This one lands. The instrument working in the negative direction is the reason to trust it in the positive one.
• Safety. 20 apparent, about three real. The rest matched “harmful” and “misuse” in passing — five of them inside the quantization cluster, which is the purest homonym this exercise produced. ADR-050 already put “safety alignment” into adaptation-and-tuning’s blurb, and the three real documents are there. Three is not several.
• Agents and tool use. 15 on the incoming list, none above two revisits, and scattered across multi-agent reinforcement learning, agentic frameworks and model reports. No subject, no engagement.
• Graphs and networks. 16 incoming, and they split. The graph-neural-network half is architecture and model-architecture covers it; the community-detection and network-curvature half is network science, and makes no claim of a kind this record’s schemes can express. The record holds two documents, one of which is a false positive.
• Federated learning and privacy. 9 incoming, none above two revisits, and six of the nine are federated learning — which is parallelism and communication, already distributed-optimization.
• Human-AI interaction (8 incoming, two or three real) and scientific domains (4 incoming). Both fail on count alone. DP-009: two is not several.

The other alternative was doing nothing, which is what the previous eleven contributions did by working down past the top of the list. That has a cost the ranking makes visible: the most-read paper on the worklist has been skipped on every pass.

Consequences

• Eighteen topics, seventeen of them primary. tiny-models remains secondary-only.
• Nine documents retagged — four to multimodal-learning as primary (LIT-071, LIT-080, SOTA-262, SOTA-271), five as a secondary. The primary moved only where the document’s central claim is about crossing modalities: SOTA-262 recommends giving each modality its own weights, and LIT-072’s open-vocabulary detection is vision work that happens to use text, so it keeps vision-and-graphics first.
• model-architecture loses a clause and gains a boundary. No document loses a tag.
• Three stale counts corrected. The vocabulary comment and blurb said “sixteen”, which was right; a third comment said the primary group was “the thirteen above” when it had been fifteen for some time. That is the second count in this file to have drifted — ADR-050 found “thirteen” where fourteen terms were declared — and suggests the number should not be written in prose at all.
• deployment-and-society is on probation by construction. It is the one term here justified entirely by what is queued rather than by what is held.