ADR-048: An unbound relation is one of four things, and liberal tagging answers one of them

Context

ADR-036 sorted twenty-four unbound relations into four kinds, retagged fifteen documents that were mis-filed on their own terms, removed two spurious sibling links, and left twenty-two standing with reasons. That classification was right and is not in question here.

One of its verdicts was reached under a filing policy that has since changed. Its fourth category ends with trunk papers whose contribution spans two topics — GShard’s conditional computation and automatic sharding, Mamba-2 as model family against the delta-rule papers as attention variants, DeepSeekMath against PPO — and the reason it left them alone was:

Splitting a trunk paper’s topic would be filing it under half its contribution.

That is correct, and it is an argument against splitting. It was written when a document carried one topic unless a second was unavoidable, so splitting was the only move available. ADR-046 changed that: tag the subject, liberally, and a paper that genuinely does two things carries two topics. The trunk-paper subcategory does not need a verdict any more — it needs the second tag it always deserved.

The report before this pass: 22 unbound relations, 10 unbound lines.

Decision

ADR-036’s four categories stand. Its trunk-paper subcategory is dissolved, and six documents carry the tags they always warranted.

Document	gains	because
LIT-187 GShard	model-architecture	the title names conditional computation
LIT-162 Mamba-2	attention-techniques	its claim is that SSMs and attention are one object
LIT-137 Gated Delta Networks	model-architecture	the gated delta rule is a new recurrence
LIT-127 DeepSeekMath	training-optimization, data-pipeline	introduced GRPO; built a 120B-token corpus by engineered selection
LIT-200 PowLU	model-architecture	an activation function is an architectural component
SOTA-032 pre-LN	model-architecture	layer-norm placement is an architectural choice

Result: 16 unbound relations, 8 unbound lines.

The test applied to each was ADR-046’s: would this tag go on if the relation did not exist? Three candidates failed it and are recorded below, because a retag pass that reports only its additions is not auditable.

Alternatives considered

• Retag the seven analysis-and-evaluation crossings. The largest remaining block, and the most tempting: a paper that measures ES fine-tuning is arguably about fine-tuning. Rejected, because ADR-036 already decided it — the invariant is over-strong where one end names a kind of work rather than a subject, LIT-230’s body argues its own filing and the argument is right, and the fix belongs to the chain declaration. Retagging exactly the documents whose edges would bind is the pattern ADR-035 forbids, and it would be indistinguishable from this pass’s legitimate half.
• Add systems-optimization to SOTA-085 and SOTA-106, binding the flash-attention edges to SOTA-083. The tag is defensible — that document’s own body is about kernels, on-chip memory per SM and tile sizes. Rejected on what the record already says: SOTA-085’s frontmatter records that a flash-attention tag was removed for binding this very edge, with the note that “the edge is real, it crosses a fault line in the vocabulary, and it should keep showing up until someone answers it.” The answer it is waiting for is a specializes: relation — SOTA-085 is an instance of SOTA-083 — and that is a scheme change. Liberal tagging does not dissolve this category and must not be used to silence it.
• Add training-optimization to SOTA-031 (micro-batch size per GPU), binding it to SOTA-092 and SOTA-093. Rejected: that practice’s claim is about memory and throughput per device, not about training dynamics, and the pair it crosses is about sample efficiency. The crossing is real — the tag would not be.
• Amend ADR-036 in place rather than supersede it. Rejected: the fifteen retags and two link removals it records are a completed pass with its own reasoning, and editing the verdict out of it would lose the fact that the verdict was right when made. DP-003 — keep what you stopped believing.

Consequences

Six documents gained a tag and appear on more topic pages, which is the point. None lost one, and no primary topic moved — every addition is second or later, so primary_topic is unchanged throughout and the indexes that read tags[0] are unaffected.

Sixteen unbound relations remain and all are wanted. Seven are the analysis-and-evaluation crossings waiting on the chain declaration, one has its far end in the attic, and the rest are the general-and-instance and rival-fixes-at-different-layers shapes ADR-036 named. The report is doing its job by continuing to show them.

The specializes: relation is now the single highest-value scheme change this report is asking for, and it is asking twice — the flash-attention pairs above, and SOTA-089 against SOTA-107. ADR-036 named it first; this decision has nothing to add except that liberal tagging did not make it go away.