ADR-010: A practice names every source it rests on, not one

Context

SOTA required exactly one source:, naming one LIT note. That was the shape ADR-001 established and DP-001 is built on: a recommendation with no paper behind it fails the build.

The rule works. The cardinality does not. Practices in this record routinely rest on several papers, and the schema could only ever record one of them. SOTA-132 is the clearest case — the 3:1 interleave of linear and global attention. Its body cites seven notes: the Kimi Linear paper that ran the controlled comparison, the two Qwen releases that shipped the same ratio independently, the Gated DeltaNet paper the linear layer comes from, Kimi K3 and Falcon-H1 for contrast. Its source: named one.

So the field was decorative for exactly the practices with the most evidence behind them, and the note bodies carried the real account. That is DP-002’s failure — a field that has stopped carrying information — with a twist that makes it worse than blankness: the field is not empty, it is confidently wrong. A reader querying “what supports SOTA-132” gets one paper and no signal that six more exist.

The cost came due this week. 39 practices — a little over a quarter of the registry — were sourced to synthetic notes that named no real paper. Every one of them looked perfectly well-formed, because a well-formed practice is one with exactly one source, and they each had exactly one. The schema could not represent “this rests on a single unreplicated claim” as distinct from “this rests on four independent groups”, so nothing could notice that a quarter of the registry was in the first category.

There is also a rule this project already wrote down and could not enforce.

LIT-121 says SOTA-122 “stays Proposed until a result from outside the authors’ group is in the record”, and LIT-134 said the same of SOTA-133 — which has since been promoted on exactly that basis, making the point. That is a countable condition — how many independent sources does this practice have? — expressed in prose because the schema offers nowhere to count.

Decision

source: takes a list of LIT codes. One entry is still the common case and still the minimum; the lint’s requirement is unchanged in force — at least one — and changed only in shape.

source:
- [LIT-133](../literature.d/LIT-133.md)
- [LIT-136](../literature.d/LIT-136.md)
- [LIT-135](../literature.d/LIT-135.md)

The order is meaningful: the first entry is the primary source, the work the practice is drawn from, and the rest are what corroborates or extends it. Generated views cite the primary and count the rest, so the index stays readable while the evidence stays complete.

What this buys, in order of how soon it pays:

• A practice’s support is countable. “How many independent sources does this rest on?” becomes a query rather than a reading. The promotion rule above becomes checkable: a Proposed practice with sources from two unrelated groups is a candidate for Active, and one that has sat at a single source for a year is a candidate for the opposite.
• A single-source practice is visible as such. Not a finding — plenty of good practice rests on one excellent paper — but a fact a curator can sort on. Had it been visible, 39 practices sharing eleven fabricated sources would have looked strange much earlier.
• The prose stops being the only true account. The seven notes in SOTA-132’s body were already the real answer; now the field agrees with them.

Alternatives considered

Keep one source: and rely on the body. This is the status quo, and its defect is not hypothetical: it is the exact configuration in which a quarter of the registry rested on nothing while every document passed the lint. A convention that lives only in prose cannot be counted, and the whole argument of DP-001 is that evidence which cannot be counted decays.

One source: plus a separate corroborated_by:. Attractive — it keeps the primary unambiguous and reads well. Rejected because it asks the filer to grade evidence at filing time, and the grade is the part they are least equipped to make: whether a second paper corroborates or merely touches the practice is a judgement that often becomes clear only later, when a third arrives. A list with a meaningful first entry captures the one distinction that is reliably knowable — which paper this came from — and leaves the rest flat. The graded version can be layered on later if flatness proves too coarse; going the other way would mean re-litigating every entry.

A weight or confidence per source. Rejected as the same problem with more decimal places. This record’s instrument for confidence is status:, and adding a second one invites them to disagree.

Consequences

Every existing practice migrates mechanically: a scalar becomes a single-element list, which changes no meaning and no rendering. The work that follows is curation rather than migration — going back through practices whose bodies already name several notes and lifting those into the field. SOTA-132 is the first.

This does not weaken ADR-001. A practice with an empty source: still fails the build; the field group’s at-least-one semantics apply to the list’s length rather than to which of several fields is present.