ADR-034: A document that changes scheme is Superseded where it leaves, and names its successor across the boundary

Context

ADR-031 created the THEORY scheme and, in the same breath, declined to use it on the documents that motivated it:

The eight finding-shaped practices stay where they are, for now. Moving one is not a status change, and “retire by changing status, never by deleting” does not cover it: a practice re-filed as a theory leaves a SOTA code that is neither Rejected nor Superseded but simply somewhere else. That needs its own decision about what the old code says and what happens to links into it, and it should be taken with a curation pass rather than bundled into the scheme that makes it possible. SOTA-010 and SOTA-012 are the clearest candidates; both are sourced to LIT-014 and both are readings of a figure.

This is that decision, taken on the first of the two.

The constraint that makes it non-obvious is that Rejected means something different in each scheme, and the difference is dangerous here. On a practice it means do not do this. SOTA-010’s claim is that skip connections smooth the loss surface, and it is true, believed, and underneath five other practices. A reader landing on SOTA-010: Rejected would read “do not use skip connections”, which is the worst sentence this record could accidentally publish.

The sota-statuses vocabulary lists “too theoretical” as a reason for the attic, which reads like an invitation to use Rejected here. It is not: that phrasing is about a paper being too theoretical to carry, and the attic blurb it sits in belongs to lit-statuses. A practice whose problem is that it was never a practice is a different thing from a practice nobody should follow.

Decision

A document that moves to another scheme is Superseded in the scheme it leaves. Three parts:

1. Superseded, not Rejected. Its blurb is “something newer does this better, and the record says what” — which is exactly true, and which carries no instruction to the reader about the subject matter. Nobody reading Superseded concludes that the underlying claim is false.

2. The successor is named with superseded_by:, which already crosses schemes. This decision was drafted asserting that no typed field could carry a cross-scheme successor, and that the successor would have to be named in prose. The lint said otherwise on the first run: superseded_by names a document in any scheme (LU-ADR-071), which is exactly the shape a move wants — typed, walked, and checked. So SOTA-010 carries superseded_by: THEORY-011 and nothing here relies on prose.

That is worth leaving in the record rather than quietly correcting, because it is the second time in this contribution that a check knew something about the record that the person reading the record did not.

3. The vacated document keeps its code and loses its body. The code is what links resolve against, and links into it exist. What it keeps is a citation line and a redirect; what it loses is the argument, which now has one home rather than two. A document that says the same thing in two places drifts — this record has paid for that twice, at #113 and at the eight themselves.

Applied to SOTA-010

SOTA-010: skip connections promote training stability by smoothing out the loss landscape is Superseded, pointing at the new THEORY document, which carries the claim, the figure’s limits, and explains: the five practices that rest on it — SOTA-032, SOTA-051, SOTA-060, SOTA-136 and SOTA-169.

No practice is filed alongside it. The obvious candidate — use skip connections — is trunk: every model in this record has them, there is nobody to tell and no live alternative to compare against, and DP-007 says the agreed trunk goes unfiled. This is worth stating because “split it into a practice and a theory” is the natural first move, and the practice half would have been an instruction with no reader.

The two compared_against: links to SOTA-011 and SOTA-012 stay on the vacated code rather than moving. THEORY has no sibling relation, and those links are a fact about how three readings of one figure were filed together — which is history, and history belongs with the document that has the history.

Consequences

Seven of the eight are still where they were. SOTA-012 — sharpness in the loss landscape correlates with test error — is the other one ADR-031 called clearest, is from the same paper and the same figure, and is not moved here. Deliberately: this decision has been applied once, and a rule that has moved one document is a rule with one data point. The next pass over the eight should take SOTA-012 first and will have a rule to test rather than to invent.

It does not satisfy ADR-031’s promotion condition, and does not claim to. That condition asks for a practice and the explanation under it moving in opposite directions, filed without either being rewritten to fit the other. This is a document moving sideways, which is a different event. ADR-031 stays Proposed.

A Superseded practice in the registry now has two meanings, distinguished only by its status note: replaced by a better recommendation, or moved to another scheme. That is a real cost and the alternative was worse. If the count grows past a handful, the honest fix is a status word of its own rather than a convention.

Alternatives considered

Rejected. Reads as do not do this on a practice. See above; it is the whole reason this decision is not a one-liner.

Delete SOTA-010 and let the links break. The record’s rule against it exists for links, and there are links.

Leave it a practice and rewrite the title as an instruction. Tempting for several of the eight and wrong for this one: the instruction would be use skip connections, which is trunk, so the rewrite would produce a practice that exists only to hold an explanation. That is the arrangement ADR-031 created the THEORY scheme to end.

Add a typed moved_to: reference crossing schemes. The right answer if this happens often. It is a scheme change to justify on one instance, and prose plus a resolvable code costs nothing until then.