Context
The LIT template said published: is “the arXiv posting month, from the id:
2205.11487 → 2022-05”, written as the first of that month. So a paper posted
on 23 October was recorded as posted on 1 October: not a coarser truth but a
false date that looks exact, and one that practices and theories then inherit
through derive: '{published}'. Every arXiv abstract page states the v1 day.
The owner’s ruling, when a reading surfaced it in the sibling record
nucleation: “that’s a stupid convention, use exact dates where we have them.”
Nucleation adopted it as its ADR-002; this is the same decision here.
Decision
published: is the exact date of first appearance:
- arXiv: the v1 “Submitted on” date from the abstract page — even when it falls in the month before the id’s (the id is the announcement month).
- DOI without arXiv: the earliest full year-month-day date Crossref reports among posted, published-online, published-print and issued.
- Neither: the date the document itself carries.
- No day anywhere: the first of the month, as before.
Two guards, because an automated source can be wrong in a way that looks exact: a Crossref date is used only when it falls in the year already recorded (Crossref’s earliest full date for an old work is often the date the publisher digitised it — a 1991 volume came back as 2006), and an arXiv v1 more than a year after the recorded date is treated as a late posting of an older work, not its first appearance. Both are flagged for a human rather than applied.
Rejected alternatives
- Keep the month convention — it records a date nobody meant.
- Trust Crossref’s earliest date unconditionally — it would move five works by years, including a 2024 preprint to its 2026 journal date.
- A
published_precision:field — nearly every LIT has a day, so it would be a field almost every record shares (DP-002).
Consequences
Backfilled on 2026-09-25 by script (arXiv abstract pages, then Crossref, both
cached): 420 LITs changed, 229 already exact or without a day-level source.
Four moved month, all arXiv v1 dates earlier than their id month (LIT-126,
LIT-139, LIT-154, LIT-515); LIT-139’s arXiv page gives 26 April for id
2606.19348, which is what arXiv states and is recorded as such. Five were
flagged and left alone: LIT-248 and LIT-296 (digitisation dates), LIT-410 and
LIT-509 (a journal issue date a year after the recorded date), and LIT-696
(the 2026 Nature date of a 2024 preprint). Every practice and theory that
derives published from its source moved with it. The template’s comment
now states the rule.