ADR-042: A practice may be filed in order to say do not do this

Context

#88 asked whether the record should accept failed ideas — “worst case scenario we end up documenting why a particular practice shouldn’t be adopted. that’s still valuable information” — and proposed two admission criteria: historical importance, and ideas that drew attention and went nowhere, with ReLoRA named as the case.

It looks like the record already does this. It does not, and the difference is the whole decision.

Eighteen practices carry Rejected today, and all eighteen mean one thing: the source does not support the claim. They are migration provenance failures, not techniques anyone tried and abandoned:

	why it was rejected
SOTA-041	states the opposite of its source — Kaplan reports larger models need a smaller rate
SOTA-064	the cited paper does not contain the word “warmup”
SOTA-101	the source performs no filtering; the string “filter” does not appear in it
SOTA-107	every occurrence of 128 in the source is a head dimension or block size, never a sequence length

Not one of them is “this was recommended, people did it, and it did not hold.” That category is empty.

The reading list does not have this gap. A LIT note is retired on the merits routinely — LIT-044 for “needs per-case tuning, no consistent cross-domain benefit”, LIT-041 for “the result does not look interesting enough to carry”. So the record already judges papers by whether they hold up, and has never judged a practice that way.

The case that is actually stuck

LIT-104 is ReLoRA, #88’s own example, and its note says in as many words that this decision is what blocks it:

Audited by #121; deliberately still unsourced. […] It is unfiled because the paper’s own status is Proposed: the record has not settled whether ReLoRA-style restarts belong in it at all, and filing the restart discipline would answer that question sideways.

So this is not a hypothetical about scope. A reading was done, a strong and specific finding came out of it — restarting an adapter mid-training requires resetting the optimizer state and re-warming the rate, and the pair without the re-warm diverges — and it is sitting unfiled waiting for #88.

Decision

A practice may be filed whose recommendation is negative. Rejected on a practice already means do not do this, so no new machinery is needed; what was missing was permission to file one that was never Active here first.

Three conditions, and only the third is new.

1. Somebody recommended it, in a paper this record holds. Already the rule — every practice names a source:, and DP-001 makes the citation structural. It does the bounding work #88 worried about on its own: a technique nobody published a recommendation for needs no counter-practice, because there is no claim to contradict. “Hyped” is not an entry criterion and does not need to be one.

2. Superseded wins wherever there is a successor. Warm restarts did not fail; they lost to single-cycle cosine, and LIT-042 → LIT-035 → WSD is a chain a reader can follow. Rejected is for a claim with no successor — the thing does not work, and nothing replaced it because there was nothing there to replace. Reaching for Rejected where a chain exists throws away the contrast DP-003 says is most of the content.

3. The failure needs a citation exactly as a recommendation does. This is the new requirement and it is the one that keeps the category honest. DP-001 says a recommendation without a citation is an opinion; a refutation without one is the same opinion wearing a status field. In practice this means a negative practice usually names two sources — the paper that recommended it and the paper that found it does not hold — and where the same paper does both, as ReLoRA’s own Table 6 does, that is the strongest case rather than a special one.

Where only the recommendation has a citation and the failure is folklore, the record says nothing. “Everyone knows this didn’t work” is precisely the claim DP-007 says has no author, and the answer there is to find the paper, not to file the rumour.

Alternatives considered

• Keep the status quo. Costs the case that is already stuck: ReLoRA has been read, has a specific finding, and its note names this question as the blocker. Doing nothing leaves a completed reading unfiled indefinitely, and leaves a reader who asks “should I restart adapters mid-training” with no answer in a record that has one.
• A new status — Refuted, Failed, or similar. Rejected twice over. The status vocabulary is closed and read by the whole toolchain (active, retires_on, the status pages, the chain walk), so a new word is a schema change; and what it would express — “never supported” versus “supported and later overturned” — is already written in status_note: on all eighteen, in prose, specific to each. A field carrying the same distinction uniformly would be DP-002’s comment with a schema.
• File these as THEORY instead, on the reading that “X does not work” is a claim about what is true. It is, but it is not what the document is for: the test is whether it tells the reader what to do, and a warning does — it says don’t. A reader deciding whether to use ReLoRA looks in the practice registry, and putting the answer in the explanation scheme hides it from the person who needs it.
• Admit anything that drew attention and went nowhere, which is #88’s second suggested criterion. Unbounded in a way the source rule is not, and it measures the wrong thing — attention is a fact about the field, and DP-005 already separates who does a thing from whether it works. Condition 1 admits every case this criterion would want and stops at the ones it should.

Consequences

The empty category is now fillable, and ReLoRA is the first candidate: its reading is done and the finding is specific. Whether it clears condition 3 is a judgement for whoever files it — the paper measures its own headline contribution as a small fraction of the gain, which is a refutation with a citation, and the citation is the authors’.

Rejected on a practice now covers two situations, and the record should expect to be asked which one a given document means. Today the answer is always “the source never supported it”; after this it may be “the source supported it and the claim did not hold”. status_note: is where that lives, as it already does on all eighteen — this decision adds a reason to keep writing those notes specifically rather than to add a field.

What this does not license: retiring a practice because it feels dated, or because adoption moved. DP-005 covers the second and consensus: is where it belongs. A negative practice is a cited claim that something does not work, and without the citation it is not a practice at all.