Context
ADR-029 added introduced_by: and made it optional. ADR-030 made it
required, and its reason was not that the values were interesting — the pass
it ran found three practices in 208 whose origin differed from their primary
source, which is the duplication ADR-029 had warned about. The reason was
what an absent field meant:
an absent field was ambiguous between “the origin is the primary source” and “nobody checked”, and those are different claims
That argument is correct and this decision does not disturb it. What it missed is that there is a third thing an author might want to say, and it is not either of those two:
the record looked, and no document it can name states this recommendation.
SOTA-311 is the case that forced it. The recommendation is to evaluate a search on tasks held out of its own fitness function. Its only source, LIT-493, is an instance of the failure: a search that did not hold anything out, and that recommends nothing of the kind. The general argument under the practice — selecting on a statistic and then reporting it inflates the comparison — is older than this corpus and standard in statistics, and the record holds no paper that states or quantifies it.
Under ADR-030 the only legal filings were both false. Writing
introduced_by: [LIT-493] asserts that Genesys first made this
recommendation, which it did not. Declining to file the practice asserts that
the record has nothing to say, when what it has is a recommendation whose
origin it cannot name — and the owner’s instruction was to file it anyway,
on the grounds that knowing the practice exists is worth more than knowing
who said it first.
Decision
schemes.SOTA.references.introduced_by.required goes to false, and the
convention around it is narrower than the schema:
- Write the field, empty.
introduced_by: [], not an omitted line. The lint reads the two identically; a human does not. An author who omits the line has probably forgotten, and an author who writes it empty has decided. - Put a comment beside it saying what was searched and did not turn up. The emptiness is an assertion — no such document exists that this record can name — and the comment is what makes it a falsifiable one. Naming the paper is how a reader refutes it.
- Say the same thing in the body. A practice with no origin should
explain, in its Source section, why its
source:is not also its origin. SOTA-311 does this in its first paragraph.
This is not a return to ADR-029’s optionality, and the difference is the
corpus rather than the config. When ADR-029 made the field optional, no
practice carried it and an absent field meant nothing at all. Every practice
carries it now — ADR-030’s pass is what put them there — so an empty one
stands out against 300-odd populated ones. The population is doing the work
the required: flag used to do.
What this costs
The lint can no longer catch a practice that simply forgot. This is a real loss and it is the price. ADR-030 bought a guarantee, and this trades the guarantee for the ability to say a true thing that the guarantee made unsayable. The mitigation is a convention, and conventions decay.
What would buy both back is an upstream change: a required: that is
satisfied by an explicitly empty list but not by an absent key — the
distinction luria currently collapses, which is why this decision had to
choose between them at all. luria lint reports an empty introduced_by: as
“no introduced_by: in frontmatter”, and those are different documents. If
that lands, this decision reverts to required: true and keeps every rule
above; the convention becomes the schema and nothing else changes. Worth
raising against luria, and not raisable from this repository.
Rejected
Fill introduced_by: with source[0] anyway and explain in prose. This
is what ADR-030 leaves as the only option, and it is exactly the failure
ADR-029 predicted for the field — reflexive duplication — except worse,
because here the duplication is not merely uninformative but wrong. It would
put “Genesys first recommended holding tasks out of the fitness function” into
a machine-readable field, where the index reads it and no prose can reach.
Mint a sentinel LIT document meaning “no known origin”. The obvious hack,
and it fails on what LIT is: a paper’s standing in the anthology. A sentinel
is not a paper, it would appear in the reading list and the lineage report,
and every practice pointing at it would declare a relation to a document with
no content. ADR-035 §4 forbids inventing a label to satisfy a check, and
this is the same move one scheme over.
Decline to file the practice. Defensible, and it was this record’s first answer — #241 exists because of it. What changed is a judgement about what the anthology is for: a recommendation that is right, cheap, and uncited is still worth a page, provided the page says which of those three it is short of. The alternative leaves the shelf empty and the reader with nothing, which is not more honest — it is only quieter.
Scope
One practice uses this today. The field stays populated everywhere else, and
a second empty one should be argued for on its own terms rather than by
pointing here. If empty introduced_by: becomes common, that is evidence the
corpus is filing recommendations it cannot source, and the response is to look
at the practices rather than at this decision.