ADR-022: The primary topic is derived from the first tag, and the lines assert it

Context

ADR-003 gave each scheme a topic vocabulary and required exactly one member per document. That is a constraint on the tags: list, and it was all luria could express: the record could guarantee a primary topic existed, but could not name it. Anything wanting to compare topics across documents had to recover it by intersecting a list against a vocabulary.

The cost came due when the chains declared an invariant (#101). What a chain asserts is that the documents in a line are the same kind of thing, and the only field available to say that with was tags — which is satisfied by any shared tag. Two practices are then “bound” by an incidental label, which is a weaker claim than the one intended and, at one edge, the wrong one.

luria 0.13 can derive a field from a template over the frontmatter (LU-ADR-089), so primary_topic can be a readable value rather than an inference.

Decision

Both schemes derive it from the first tag:

[luria.schemes.SOTA.fields.primary_topic]
derive = "{tags[0]}"

and both chains assert that instead of tags:

invariant = "primary_topic"

The tag group stays exactly as it was. The group guarantees that exactly one tag comes from the vocabulary; the derivation says which tag that is. They answer different questions and neither replaces the other.

Derived rather than stored, so the value cannot drift: retagging a practice retags its primary topic in the same edit, and there is no second place for the fact to live and go stale. Nothing is written to any document — 406 files gain a field and none of them changes.

What the stronger invariant re-opens. Exactly one edge, SOTA-085 / SOTA-161: using flash attention, filed under attention-techniques, and keeping its output in FP32 because the rounding bias compounds, filed under model-stability. #101 bound that pair by giving both a flash-attention secondary tag — the only move available, since exactly-one forbids a second primary — and doing so satisfied the check rather than answering it. Their primary topics genuinely differ, the edge genuinely crosses a fault line in the vocabulary, and that is what these findings are for. It joins the twelve in the reading list on #85’s worklist. The flash-attention tags stay: they are true, and they are useful for browsing whatever the chain asserts.

On the reading list the change reports the same twelve edges either way, because no note carries a secondary tag today. It is declared there anyway, so both chains assert the same thing for the same reason, and because a note will carry a secondary tag eventually.

Alternatives considered

• Keep invariant = "tags". The status quo. It is not a weaker version of the same check, it is a different claim — “these share a label” rather than “these are the same kind of work” — and a record that adds a tag to bind an edge is answering the check rather than the question. #101 said as much at the time; it had no other option.
• Store primary_topic: as an ordinary field on every document. Explicit, greppable, and immediately a second copy of a fact the tags: list already carries. Retagging a practice would then have to remember to update it, and the two would disagree the first time someone forgot — the failure this record has already had once, with the topics that sat unapplied in a Jinja template until ADR-003.
• A separate primary_topic vocabulary, with the tag list moved out of the group. Would let the derived value be checked against a controlled vocabulary, closing the hole below. It also splits one vocabulary into two that must agree, which is the same duplication in a different place.
• Wait for the guard. Defensible: the convention below is unenforced, so this ships a silent failure mode. Rejected because the convention already holds everywhere, the alternative is a known wrong invariant rather than a possible wrong tag order, and the guard is filed rather than assumed.

Consequences

A convention nothing checks: the primary tag must be written first. The tag group guarantees exactly one tag from the vocabulary; it says nothing about position. Write the tags in the other order and primary_topic derives to a secondary tag — the group is still satisfied, the derivation still renders, and the field is quietly wrong, along with the chain invariant that reads it. It holds in all 188 practices and all 218 notes today; it is a silent error the day it does not. Filed upstream as LU-#229, which is where the check belongs — the violation is always wrong and the fix is mechanical.

primary_topic is now a field, so it can be faceted on, reported, and read in a template. That is the payoff beyond the invariant, and it is worth saying out loud: the topic consolidation stopped being only a rule the lint enforces and became a value the record can use.