Context
LIT-425 is the T5 paper. It entered this record because forty-six documents
train on C4 and the corpus had no document behind it, so it was filed with
one tag: data-pipeline. The reasoning written down at the time was that the
record holds the paper for its corpus, and that a second tag would be
incidental and might bind a chain edge under ADR-035.
Both halves of that were wrong, and they were wrong in ways worth separating.
The subject was mistaken for the motive. T5’s text-to-text framework and
its systematic ablation of transfer-learning choices are not incidental to
the paper — they are most of it. Why the record went and got a document is a
fact about the record; what the document is about is a fact about the
document. Only the second belongs in tags:, because tags: is what a
reader browses by, and a reader looking for architecture work has no way to
know that this project’s interest was the dataset.
The chain warning was applied far outside its case. ADR-035 says not to
add a tag in order to bind a relation, and the guard exists because
chains.lineage declares invariant: tags. But an invariant binds an edge
only where an edge exists — where a document declares extends, corrects
or compared_against. LIT-425 declares none. There was no edge for the tag
to bind and never could have been.
The cost of the error is asymmetric and that is what settles it. An extra
true tag puts a document on one more index page. A missing one makes the
document unfindable to everyone who browses that topic, silently, with
nothing in any report to say so — DP-004’s shape again, where the absence
leaves no trace.
Decision
Tag the document’s subject, and be liberal about it. More than one topic is normal; three is fine. The test is whether the tag is justifiably appropriate — whether someone browsing that topic would be right to expect this document — not whether it is the single best word, and not whether it is why the record acquired the thing.
ADR-035 stands unchanged: the first tag is still the primary topic, and
adding a tag in order to bind a relation is still forbidden. What this
decision says is that the forbidden case is narrow. A tag that is true of the
document is not made suspect by a chain that might one day read it, and a
document with no relations declared is not in the chain’s reach at all.
Alternatives considered
- Minimal tagging — one topic unless a second is unavoidable. What was being practised. It optimises for a tidy index at the cost of the thing an index is for, and it makes every additional tag an argument to be won rather than a fact to be recorded.
- Tag by filing motive, so the tags record why the record holds a
document. Coherent, and wrong for the reader: it encodes the curator’s
history where the browser expects the subject. The motive already has a
home — the
Standing in the anthologysection says it in prose, which is where a reason with nuance belongs. - Drop
invariant: tagsfromchains.lineageso the tension cannot recur. Too big a change for the problem: the invariant does useful work on documents that do declare relations, and the fix here is to read its scope correctly rather than remove it.
Consequences
LIT-425 v2 gains model-architecture and adaptation-and-tuning;
LIT-426 v2 gains analysis-and-evaluation. Both were filed under the
old reading in the contribution immediately before this one.
Documents already in the record are likely under-tagged, since the cautious reading has been in force for a while. This decision does not mandate a sweep — a retag pass is its own contribution, and the topic pages are generated, so the cost of being late is a document that shows up on one fewer page until someone touches it.
The chain invariant now has a stated scope, which it lacked: it binds where a relation is declared and nowhere else. A future change that makes tags bind edges more broadly would need to say so, because this decision reads them as not doing that.