ADR-036: An unbound relation is one of four things, and only one of them is a retag

Context

Both chains in this record assert an invariant on primary_topic: two documents joined by succession or rivalry are about the same kind of thing, and a topic in common is what says so. docs/reports/unbound-lineage.md reports where the assertion was made and nothing says what it means. It had 38 unbound relations and 16 unbound lines, and had never been read through.

The report’s own preamble gives the prior — “usually a sign the vocabulary is short a word rather than that the relation is wrong” — and luria.yaml records the failure mode beside the practice chain: #101 bound the SOTA-085 / SOTA-161 pair by giving both a flash-attention secondary tag, which satisfied the check rather than answering it. Any pass over this report has to be able to leave a finding standing.

Reading all 38: they are four different things.

Decision

1. Fifteen documents were mis-filed on their own terms. Retagged.

Not to bind an edge — each is wrong read alone, and would still be wrong if the relation did not exist.

• Eight positional-encoding papers — LIT-045 (RoPE), LIT-048 (ALiBi), LIT-192 (Position Interpolation), LIT-193 (YaRN), LIT-207, LIT-208, LIT-209, LIT-210 — split between model-architecture and attention-techniques, because when they were filed there was no third option. representation-and-encoding names positional encoding in its own blurb, and “the frequency or basis choices that go with them”, which is literally YaRN’s subject and LIT-210’s. #114 added that topic and nobody went back. All eight now carry it, and NOTE-028 follows its paper.
• LIT-211, Evolution Strategies at Scale: LLM Fine-Tuning Beyond Reinforcement Learning, carried training-optimization. It is about fine-tuning; SOTA-154, the practice it introduced, and five of its six descendants already carry adaptation-and-tuning.
• SOTA-092, SOTA-093, SOTA-094 — three claims about batch size from one model report, carrying model-architecture because a model report is where they were found rather than what they are about. training-optimization names batch size.
• SOTA-060, an initialization rule for stability, carrying distributed-optimization because its source is a Megatron paper.
• SOTA-098 and SOTA-099, both about detecting instability — SOTA-099’s title says so — carrying training-optimization while the practices they are compared against, one of them from the same paper, carry model-stability.

This is the largest category and it is a filing error, not a vocabulary one: eleven of the fifteen were filed under the topic of the document they were found in rather than the topic of the claim they make, which is exactly the distinction ADR-026’s filing rule draws.

2. Two sibling links were analogies, not comparisons. Removed.

ADR-011 defines compared_against: as this work evaluates itself against that one. Nobody compared SOTA-045 (profile the input pipeline separately) with SOTA-091 (profile memory access patterns), or SOTA-047 (overlap communication with the backward pass) with SOTA-079 (pre-fetch the next batch). Different papers, different subsystems; the shared thing is a word in the title and a shape of idea. Both links are removed, with the reason left in the frontmatter of all four practices.

This is the case the report’s preamble calls unusual, and it is worth naming that it exists: an unbound relation is sometimes evidence that the relation was asserted too cheaply.

3. Seven relations cross because analysis-and-evaluation is a kind, not a subject. Left standing.

LIT-127 and LIT-211 against LIT-230, LIT-233 and LIT-235; LIT-235 against LIT-238. Every one joins a paper that does something to a paper that measures it. They are about the same thing at different altitudes, which is precisely what the relation asserts and what the invariant cannot see.

Twelve of the thirteen topics name a subject. analysis-and-evaluation names a kind of work — “how to find out whether something worked… theory, interpretability and debugging belong here too” — and it will unbind a lineage every time a line grows an analysis paper. This is not a defect in those six documents’ tags: LIT-230’s body argues its own filing explicitly, and the argument is right.

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, and it belongs to whoever next opens the chain declaration. Nothing here is retagged to hide it.

4. The remaining fifteen are real fault lines. Left standing, and named.

Four shapes, and they recur:

• A general practice and its domain-specific instance. SOTA-083 (custom kernels for critical ops) against SOTA-085 and SOTA-106 (flash attention): in both pairs the second is an instance of the first. The record has no relation for that, so compared_against is standing in — which is why these read as comparisons nobody ran while not being spurious in the way category 2 was. A specializes: relation is the thing that would answer these, and it is a scheme change, not a retag.

• One edge whose far end is in the attic. SOTA-089 (align tensor dimensions to hardware boundaries) against SOTA-107, which is Rejected: its status note records that the 128 in it was traced to a head dimension and a block size, never a sequence length. That looks like the shape above and is not one. An unbound relation with a retired end is not a filing question — the comparison is the record of how the claim was checked, which is what the attic is for, and this report does not filter by status. Worth knowing when reading it: some share of any future finding here will be this, and this pass found one in twenty-four.
• Rival fixes made at different layers. SOTA-032 (pre-LN) against SOTA-100 (warmup proportional to model size) — an architectural fix and a schedule fix for one instability. SOTA-131 (rescale query and key weights under Muon) against SOTA-192 (normalize queries and keys) — an optimizer-specific fix and an architectural one. SOTA-085 against SOTA-161 — flash attention and the FP32 output its rounding requires. LIT-030 against LIT-200 — an activation function and the one that replaces it for stability. These cross because the layer you fix something at is not the layer the problem is about, and that is a true thing about the work.
• Trunk papers whose contribution spans two topics. LIT-187 (GShard: conditional computation and automatic sharding) against the three MoE architecture papers around it. LIT-137 against LIT-162, where Mamba-2 is filed as a model family and the delta-rule papers as attention variants. LIT-211 against LIT-229 and LIT-242, where an LLM fine-tuning method meets the optimizer papers it descends from; LIT-127 against LIT-241, GRPO against PPO. Splitting a trunk paper’s topic would be filing it under half its contribution.

Consequences

The report is down from 38 relations and 16 lines to 24 and 10, and the part that remains is the part it is for. The positional-encoding line — eight documents, six relations — is gone entirely, which is the shape of a finding that was a missing retag rather than a missing word.

The vocabulary is not short a word. That was the report’s prior and on this reading it is wrong: not one of the 38 wanted a fourteenth topic. What the record is short of is a relation (specializes:), and what the chain is short of is a way to say that one end of an edge is an analysis of the other.

Eleven of fifteen mis-filings shared one cause, and it is worth stating as a filing habit rather than eleven corrections: a practice extracted from a model report, a Megatron paper or a systems paper tends to inherit that document’s topic instead of its own. ADR-026’s rule already covers it — take the topic of the claim, not of where it was found — and this is the first measurement of how often it is missed.

Two removed relations is not a licence to prune. They were removed because compared_against has a definition in ADR-011 and they did not meet it, not because they were unbound. The fifteen in category 4 are equally unbound and all of them stay.

Alternatives considered

Bind the flash-attention cluster with a secondary tag. Already tried, at #101, and luria.yaml records why it was wrong: the invariant moved to primary_topic partly to re-open that edge. Doing it again under a different name would close the finding without answering it.

Add a fourteenth topic for the analysis/subject split. It would be a second axis wearing the costume of a topic. analysis-and-evaluation is already that second axis; adding another would not stop a lineage crossing between them.

Retag the Mamba papers to attention-techniques to bind their line. The record’s practice here is inconsistent — Mamba, Mamba-2, Mamba-3 and RWKV-7 are model-architecture while Kimi Linear, Gated DeltaNet, GLA, DeltaNet, Griffin and Monarch Mixer are attention-techniques — but there is a defensible rule under the inconsistency: a model family takes model-architecture, a mechanism takes attention-techniques, and Mamba is both. Picking a side across six documents on one filer’s judgement, to bind one line, is the #101 move. It is named here so the next pass has something to argue with.

Loosen the invariant to tags rather than primary_topic. That is the declaration this record deliberately tightened, for a reason its own comment states: any shared tag is too weak to carry the assertion.