Context
The #180 pass filed papers the owner returned to on three or more days. Its readers hit the edge of the vocabulary seven times, and each one said so rather than forcing a tag:
- The Work Capacity of Channels with Memory and The Computational Foundations of Collective Intelligence were declined: one is about the work a finite-state agent can extract in a percept–action loop, the other about animal collectives, and no topic is about agents or what acting does.
- Community Detection on Networks with Ricci Flow was declined: no topic is
about graphs, though the record already holds the GNN expressivity and
evaluation papers (LIT-580, LIT-581) and a practice and a theory drawn from
them (SOTA-350, SOTA-351, THEORY-083), all filed under
model-architectureandanalysis-and-evaluation. - From monoliths to modules (LIT-759), a paper about world models, was
filed under
model-architectureand said its primary subject had no word. - Erwin (LIT-763) evaluates only on cosmology, molecular dynamics, PDEs and airflow, and its note said the domain was unplaceable.
- Textoshop (LIT-741) was filed under
deployment-and-society, a topic about deployed systems among people, with a note that the fit was loose; LIT-503 had been filed the same way.
The owner approved the four words the readers proposed.
Decision
Four topics, each with a label and blurb in luria.yaml:
agents-and-environments— systems that perceive, act and are changed by the consequences: percept–action loops, world models and training environments, multi-agent and collective behaviour.graphs-and-networks— graph-structured data and the models that run on it: message passing, how GNNs are evaluated, curvature, rewiring, community structure.physical-sciences— models whose data is the physical world: learned simulation surrogates and foundation models for physical observation. It sits besidebiomolecular-modeling(ADR-059), which keeps molecules.human-ai-interaction— how people work with a model through an interface, how such interfaces are scored, and what lab studies of them can show.deployment-and-societykeeps what a deployed system does among people at scale.
The filing rule stays ADR-026’s and ADR-059’s: a document about the subject takes the topic first; a general method found there takes its kind first and this topic as well.
Retagged. First tag:
graphs-and-networks: LIT-579, LIT-580, LIT-581, LIT-582, THEORY-083, SOTA-350, SOTA-351;agents-and-environments: LIT-759;physical-sciences: LIT-715 (Surya, a heliophysics foundation model);human-ai-interaction: LIT-741, LIT-503, THEORY-091, SOTA-309.
Secondary tag: SuperGlue (LIT-386) takes graphs-and-networks; AgentSociety
(LIT-697) and MemGraphRAG (LIT-502) take agents-and-environments; Erwin
takes physical-sciences; LIT-717 takes human-ai-interaction.
Every document keeps its existing tags, so no relation’s invariant: tags
loses a shared value.
Alternatives considered
- A separate
collective-intelligence. Its reader offered it as the alternative ifagents-and-environmentswere kept to machine agents. The blurb includes collective behaviour instead: a group that perceives and acts is an agent system, and splitting it would leave one paper alone under a topic. - Leaving the three papers out. They are already held in nucleation. ADR-059 is explicit that the topics are an axis, not a scope, and the owner asked for them to be held in both.
- Widening
deployment-and-societyfor interface studies. Its subject is a system’s effect on people at scale, and its blurb is about audits and information ecosystems; a twelve-person lab study of an editor is a different kind of evidence, and the record’s warnings about it (SOTA-298, SOTA-303) are about exactly that difference.
Consequences
- The vocabulary has twenty-six topics;
CLAUDE.mdand the vocabulary’s own header say so. - The three declined papers are filed in the same contribution.
- The GNN material now shares a topic, so a reader can find the expressivity bound, the evaluation pitfalls and the practices drawn from them together.