Context
ADR-029 added introduced_by: and made it optional, for a reason it stated
plainly: for most practices the origin is the primary source, and a field
every record shares is DP-002’s failure. It also named the risk that the
field would be filled in reflexively with source[0]. It filed itself
Proposed with a condition for promotion:
A pass over the practices that populates the field finds cases the single-list arrangement could not express — practices whose origin is not their primary source — in numbers that make the second field carry information rather than duplicate the first.
That pass has now run over all 208 practices, and the honest result is mixed:
| practices | |
|---|---|
| cite exactly one source | 168 |
| cite several; the primary is already the earliest | 22 |
| cite several; an older paper sits in their own list | 18 |
And of all 208, exactly one carries a comment naming an origin that is
not its primary source — SOTA-121, whose source block already said “LIT-159
is the origin” in prose that no field could hold. Three practices in total
now differ from source[0]: that one, plus two of the three ADR-029 filed
with it — the third names the same paper in both fields, and says so in its
own comment.
So the pass did not find the numbers the condition asked for. Three in 208 is not “carries information rather than duplicates”; it is the duplication ADR-029 warned about, arriving exactly as predicted.
Decision
introduced_by: is required on every practice. All 208 carry it; 205
name the same code as source[0], and three do not.
This overrides ADR-029’s optionality against the evidence its own promotion condition asked for, and the reason is a different one than that condition tested. The condition asked whether the field’s values would differ often enough to be worth a column. The answer is no. What makes the field worth requiring is what its absence meant:
Under required: false, a practice with no introduced_by: was ambiguous
between two very different states — the origin is the primary source and
nobody has checked. Those are not the same claim, and a record whose whole
purpose is that a recommendation can be traced to work that justifies it
should not have a silent third state. Requiring the field collapses the
ambiguity: every practice now asserts an origin, and an assertion is a thing
a future reader can find wrong. SOTA-087 was dating a 2019 recommendation to
2022 for years, and what let it sit there was that nothing had ever been
asserted for anyone to check.
The scaffold carries the field, with a comment saying the usual case is the
same code as source: and that a difference is worth a comment. That is
where the reflexive-copy risk is actually managed — at filing time, by
someone with the paper in front of them — rather than by leaving the field
off.
ADR-029 is Superseded by this decision. Its substantive holding is
carried forward unchanged: the field exists, it means the work that first
stated the recommendation, source: is not reordered, many: true for
parallel invention, and published: still derives from source[0].
Alternatives considered
- Leave it optional, as ADR-029 decided. The evidence supports this, and it is the alternative that the measurement actually argues for. Rejected because it preserves the ambiguity above: the 205 practices where the origin is the primary source would keep saying nothing, and “says nothing” is indistinguishable from “was never looked at”. The cost is real and is stated in the consequences.
- Require it only where a practice cites more than one source. This
targets the 40 practices where the question can even arise, and leaves the
168 single-source ones alone. It is the most defensible middle, and luria
can express it (
required_when). Rejected because the ambiguity it leaves is the worse half: a single-source practice is exactly the one where an unexaminedsource[0]is most likely to be wrong, as SOTA-087 was. - Populate by research rather than by default. Determining each practice’s true origin paper, as was done for the sliding window, would make all 208 values earned rather than assumed. This is the right thing and it is not a backfill: it took a curation pass and a new LIT note to do it once. The field being required is what makes that work findable — a wrong origin is now a correction to a stated field rather than a gap nobody sees.
Consequences
205 of 208 practices now assert an origin identical to their primary source, which is close to the shape DP-002 warns about, and the warning should be read rather than waved past: if a year from now the count of practices where the two differ has not moved, the field is a comment with a schema and this decision was wrong.
The lint now catches an omission at filing time. A practice scaffolded
from the form starts with the field; one that drops it fails. The scaffold
violation this change first produced — introduced_by: required but not
scaffolded — is the mechanism working, and is fixed here.
Three practices differ, and each is a correction the record could not previously state: SOTA-060 and SOTA-087 (the sliding window and the recomputation trick, both Child et al. 2019 rather than their 2021-22 evidence), and SOTA-121 (Muon, whose origin the source comment had been carrying in prose).
What this obliges: the population was mechanical, and mechanical work asserts nothing. The 205 defaults are claims now, and a wrong one is a finding rather than a silence — which is the point, but it means the record carries 205 assertions no one has individually checked. That is a debt this decision creates deliberately and should be worked down by curation, not by another pass over the frontmatter.