ADR-029: A practice's origin and its evidence are different relations

Context

DP-005 cut adoption away from evidence: who ships a thing and whether it works are different questions, so source: holds the work that produced evidence about the claim and consensus: holds who does it. That cut was correct and it was only one of the two that source: needs.

What is left inside source: is origin and evidence, still merged. The practice template says so in as many words — “the first is the primary source; the rest corroborate — a replication, the production report that shipped it, the paper that argues the mechanism” — and every item in that list is a kind of evidence. There is no slot for the paper that first said to do this.

ADR-017 then defined the primary source as the one that would force a rewrite if it were retracted. That test is right for what it was written for and it selects the paper whose argument the body uses. The paper that introduced a recommendation is frequently not that paper, because a practice’s body is usually written from the clearest or most recent treatment.

Three cases, all found while filing one 2019 paper:

• SOTA-087 — recompute attention during the backward pass instead of storing it — is sourced to FlashAttention (Dao et al., 2022). Child et al. (2019) §5.4 recommends exactly this, and for long-sequence attention specifically: “memory usage is high for these layers relative to the cost of computing them”. Dao’s contribution is that the recomputation becomes nearly free once the kernel is memory-bound — which is why the practice’s body is correctly Dao’s argument, and why reordering source: would break it.

• SOTA-060 — depth-scaled residual-output initialisation — is sourced to Megatron-LM (Narayanan et al., 2021). Its body names the actual origin and does not cite it: “the same reasoning that puts a 1/√(2·n_layers) factor on the output projections in GPT-2-style initialisations”. That factor is Child et al. (2019) §5.2, with its reason — keeping the ratio of embedding scale to residual-block scale invariant in depth. An uncited claim inside a practice, in a record whose first design principle is that the citation is structural.

• SOTA-138’s Sequence section begins at NSA (February 2025). Fixed sparse attention is where that line starts.

This is the defect DP-007 describes and the record has now hit twice: SOTA-132 was “describing a line from its middle” (2026-09-08), found by hand, by someone following a citation backwards. Nothing mechanical can find it, because a practice citing only its most recent treatment is indistinguishable from a practice whose origin is its most recent treatment — the two look identical from inside the record, and the field that would tell them apart does not exist.

Decision

Add introduced_by: to the SOTA scheme — LIT, optional, many, no converse (it crosses schemes, and luria’s converses are same-scheme).

introduced_by:
  scheme: LIT
  required: false
  many: true

source: keeps its meaning exactly: the work that produced evidence about the claim, ordered, first entry primary, published: derived from it, ADR-017’s retraction test unchanged. introduced_by: names the work that first stated the recommendation.

Three things that look like details and are load-bearing:

1. A code may appear in both, and usually will. A paper that introduces a recommendation normally also ran the experiment for it. The fields are not a partition; they answer different questions about the same citation.

2. many: true, for parallel invention. The record already holds independent arrivals at one design — LIT-208 and LIT-209 reach the same layout three months apart, and the curation journal treats that convergence as stronger evidence than either alone. An origin field that could hold only one of them would be forcing a false priority claim.

3. Not required, and deliberately. Most practices will never populate it, because for most the origin is the primary source. A field required everywhere would be filled with copies of source[0], which is DP-002’s failure — a field every record shares is a comment with a schema.

Alternatives considered

• Reorder source: so the origin comes first. The obvious one, and it was what this session started doing before the second field was proposed. It breaks two things at once: published: derives from source[0], so SOTA-087’s date would become 2019 while its body argues Dao’s 2022 mechanism; and ADR-017’s retraction test would no longer describe the field it defines. Worse, it is lossy in the direction that matters — it can record that one paper matters most, never why, so the next reader cannot tell an origin from a strongest-evidence.

• Say it in prose. The status quo, and the reason SOTA-060’s body names the 1/√(2N) factor and cites nothing: prose citations are exactly what the record moved away from, because a paragraph cannot be counted, cannot be found to cite a retired document, and cannot answer “what rests on this paper”. This is ADR-011’s argument for lineage fields, applied one relation over.

• A qualified reference — source: [{code: LIT-x, role: origin}]. Rejected for the reason ADR-011 gives for keeping the sign in the field name rather than on the reference: an edge property that requires opening a mapping to read is not legible in a listing, and luria’s reference machinery is declared per-field.

• extends: from the practice. Wrong shape. extends: relates a practice to a practice; a 2019 paper’s recommendation is not a SOTA document in this record and filing one just to be extended would invent a document to hold a citation.

• Do nothing until a pass counts the cases. Rejected on the grounds that the three known cases were found incidentally, by one person filing one paper, in one afternoon. That is not a sample that argues the number is three.

Consequences

Filed Proposed, with a condition, because three cases found by accident is evidence that the field is needed and not yet evidence of how much it carries. The promotion test is a deliberate pass — the query DP-004 asks for, run backwards from papers to the practices that cite them — not an accumulation of incidental finds.

What this obliges: the pass. introduced_by: is worth having only if somebody runs the backwards query over the practices whose primary source postdates the technique they describe, and the record currently has no view that surfaces those. Until then the field is populated where the defect happened to be found, which is honest and is not coverage.

What it costs: one more field to reason about when filing, and a real risk that it gets filled in reflexively with source[0]. The required: false is the defence, and the field is worth nothing if a future template starts scaffolding it with a placeholder.

Populated on three practices in the contribution that proposes it — SOTA-087, SOTA-060 and SOTA-138 — so that the decision ships with the evidence for it rather than an intention to find some.