ADR-049: Unbound is never the resting state: the relation is evidence that a tag is missing

Context

ADR-036 sorted twenty-four unbound relations into four kinds and left twenty-two standing with reasons. ADR-048 re-examined them, bound six, and left sixteen — reporting the largest remaining block, seven crossings where a paper that measures something meets a paper that does it, as a defect in the invariant rather than in the documents:

So the invariant is over-strong where one end is analysis-and-evaluation. Fixing that is a change to what the chain asserts, not to a tag.

That is wrong, and I then made it worse by proposing an unbound-ok: acknowledgement directive to luria (LU-#297, closed) so the crossings could be marked deliberate. Both moves treat the report as a noticeboard.

It is a defect detector. A relation asserts that two documents have something in common. If a chain exists, the commonality exists — the author could see it, which is why they wrote the relation. So an unbound relation is one of exactly two things, and both need fixing:

• the invariant is missing — nothing names what the line has in common; or
• the relation is wrong — the chain asserts a commonality that is not there.

There is no third case. invariants.py’s own docstring says so and both earlier passes read past it: “Its usual reading is not that the documents are mis-related but that the vocabulary is missing a word.”

Decision

Unbound is never the resting state. Every finding gets one of the two verdicts above. A recurring finding is not evidence that the report is over-strong; it is evidence that nobody has answered it yet.

Twenty-one documents gained a tag they were already about. All of them passed ADR-046’s test — true independent of the edge — and the relation is what pointed at them:

line	gained	on
evolution strategies for LLM post-training	adaptation-and-tuning	LIT-230, LIT-233, LIT-235, LIT-236
the same, as optimizer work	training-optimization	eight more of its sixteen members
SSM / delta-rule sequence mixing	attention-techniques	LIT-161, LIT-165
early-training instability	model-stability	SOTA-008, SOTA-009, SOTA-100
batch size	training-optimization	SOTA-031
the flash-attention kernels	systems-optimization	SOTA-085, SOTA-086, SOTA-087, SOTA-106, SOTA-107, SOTA-108, SOTA-161
query/key interventions	attention-techniques	SOTA-131, SOTA-192

Result: 0 unbound relations and 0 unbound lines, from 22 and 10.

SOTA-085’s standing request is answered. Its frontmatter recorded a tag removed for binding its edge to SOTA-083, with the note that the crossing “should keep showing up until someone answers it.” It has: flash attention is a custom kernel, systems-optimization is what that document’s own body is about, and the tag was missing rather than forbidden.

The groups.primary_topic block is removed from all three schemes. It listed thirteen of the fourteen topics under require: any — which luria documents as “the group is a label, not an axis”, constraining nothing. Its comment still claimed “the topic consolidation, enforced… exactly one primary per practice”, true before ADR-035 relaxed it from exactly-one. Once inert, its membership drifted: tiny-models was added to the vocabulary and never to the list, so the config quietly asserted that a document about small models has no primary topic. DP-002 names the shape — a field every record shares is a comment with a schema — and an inert one that has stopped being true is worse than that.

Alternatives considered

• An unbound-ok: directive. Filed upstream, then withdrawn and closed. It reasons by symmetry with inactive-ok: and the analogy fails: a citation of a retired document is a legitimately permanent situation, and an unbound relation is a defect. The directive would have let a record silence exactly the findings the report exists to raise, and the ones most likely to be silenced are the recurring ones — which is to say the ones with a real structural defect behind them.
• Drop analysis-and-evaluation from the vocabulary, since it names a kind of work where the others name subjects. Rejected on measurement: 91 documents carry it and for 74 it is the only tag — and those 74 are benchmark and evaluation notes where it is the subject. They declare no relations and produce no findings. Exactly 8 documents carried it alone and sat in a chain, and those 8 were the whole problem. The tag is fine; it was standing in for a subject in eight places.
• Change what the chain asserts — ADR-048’s proposal, and ADR-036’s before it. It would have hidden twenty-one genuinely missing tags.
• Add more kinds of work to the vocabulary, so the axis analysis-and-evaluation belongs to is properly populated. Not rejected — deferred, because nothing here needed one. The rule that makes it safe is the one this decision establishes: a kind-of-work tag never stands in for the subject of a document that is about something. Add a kind when a document wants it, as ADR-027 did.

Consequences

The report is now silent, and that is the point. Every future finding is new, and the two verdicts are the whole vocabulary for answering it. The next unbound relation means something is wrong today rather than joining a standing backlog a reader has to re-derive.

tags[0] ordering is still load-bearing and still unguarded. Removing the inert group does not change that; LU-#229 is where the guard is asked for. The reasoning that was buried in the group’s comments now sits on the tags field where it applies.

Two decisions were reversed to get here, and both were reasoned. The lesson worth carrying is narrower than “they were wrong”: each treated a recurring finding as evidence about the checker. A check that keeps firing on the same input is usually right about the input.