Context
Issue #163 listed AlphaFold among things to file. Working through the
issue, the filer left it unfiled and reported it as a scope question
because “protein structure prediction is outside the eighteen topics” (the
count at the time, before capability-thresholds). The owner’s answer is
the rule this decision writes down:
the topic set is not a constraint, it is just an axis around which to organize content. if the axis is insufficient to represent content we’re interested in: that should be interpreted as a signal about the inadequacy of the axis rather than the non-relevance of the content.
The record already said most of this. The topics alert says the
vocabulary is closed “NOT because the list is finished”. ADR-026 says the
bias toward language-model training is “a bias, not a boundary”, and
DP-009 is written beside deployment-and-society in luria.yaml:
“declining a document because no category fits is the self-confirming
reading”. The slip happened anyway, because the sentence a filer reads
first, in CLAUDE.md, said work “qualif[ies] if the recommendation is one
the nineteen topics can express”. Read literally, that sentence is a gate.
A sweep found the cost was already being paid:
LIT-505(ESM-1b, a protein language model trained on 250 million sequences) andLIT-506(RNA-FM, an RNA foundation model for structure and function) are filed withrepresentation-and-encodingas their primary topic. That is the nearest wrong word thetopicsalert warns about. Their subject is models of biological molecules, and a reader browsing for that had nowhere to look.- AlphaFold 2 and 3, which the issue asked for, had no topic at all.
Decision
A twentieth topic, primary-eligible:
biomolecular-modeling— models whose data is molecules: protein and nucleic-acid structure prediction, biological sequence language models, docking and molecular property prediction. What they get right, what their benchmarks can show, and which of their methods travel to other domains.
The filing rule stays ADR-026’s. A document about biomolecules takes this topic, first. A practice that was discovered in AlphaFold but is about something general, such as self-distillation on confident predictions, takes its kind first and this topic as well, so someone browsing either finds it.
The scope rule, stated where filers read it. CLAUDE.md now says the
topics are an axis, not a scope, and that content the axis cannot place
means the axis gains a topic. The document is not declined. The topics
blurb in luria.yaml says the same, so it renders in the views.
Retagged: LIT-505 and LIT-506 take biomolecular-modeling as their
primary topic and keep their other tags. Their subject is protein and RNA
models, and their representation and adaptation lessons stay findable
under the old words.
Rejected alternatives
- Leave AlphaFold unfiled until a transferable lesson is identified. This is what the filer proposed, and it inverts the relation between content and axis. It also misjudged the content: AlphaFold’s self-distillation and confidence heads are general methods, and would have been filed under their kinds anyway.
protein-structure-prediction. It names the issue’s item exactly and nothing else. AlphaFold 3 predicts nucleic acids, ligands and their complexes, and RNA-FM is about RNA. A word that fits its first document and not its second would need replacing on the next filing.- A general
scientific-applicationsword. It is broader than anything in hand, and a topic nobody can blurb precisely is one nobody can tag consistently. Physics, climate and materials models can each get their own word when the record holds them.
Consequences
- Twenty topics, eighteen primary-eligible.
tiny-modelsandcapability-thresholdsstay where their decisions put them. LIT-579(Where did the gap go?) uses molecular benchmarks but is about evaluation practice. It is not retagged, because someone browsing biomolecular models would not expect it.- The next document the axis cannot place follows the same path: a word, a blurb, a decision. It does not wait for permission.