Found by putting a real model at the end of the chain: a page of notes from a
theological book composed a prompt asking who wrote Matthew 4:17, and the
model rightly answered nothing for all twenty-three entries.
Those are scripture references, which is most of what the notes of such a book
are, and they are now not references at all — no record, and no question put
to anybody about them. A journal and its volume look similar on the page and
are still read as what they are.
Also: the stop that belonged to a marker no longer opens the entry after it.
Tried against three of the largest books in the corpus — 2,700 chunks, 44,000
citations the model had already extracted — and two things showed up that the
smaller sample had not.
A chunk that begins part-way through an entry leaves a closing quotation mark
with no opening one, and the title read off it starts mid-word and runs to the
end of the line. A title carrying a stray quote, an escape, or an editor's
credit is now no title, and `unsettled` says so.
The optional fields are written as null rather than left out. A key that comes
and goes is a key every consumer has to guard, and the documented Python use
raised KeyError on the first reference without a page.
Measured against the citations an LLM had already extracted from 40 books of
the same corpus — 19,340 of them over 3,424 chunks — and improved where the
comparison showed a real gap rather than a difference of task.
On pages of notes, where most citations live: year agreement with the model
44% -> 56%, title 18% -> 23%, author 16% -> 24%.
What changed:
- A note listing several works, semicolon-separated, becomes one record per
work. The note number is kept on all of them.
- A full stop is not a separator. It parts the author from the title in every
bibliography entry there is.
- A page of body text with notes at its foot no longer reads the body as the
first note.
- A run carrying a page reference or a semicolon is not reported as a title.
- Six sentences with no imprint, no page and no identifier is a remark about
the argument, not a reference.
- A translator standing where the journal goes is a credit, not a container,
and `(London, 1973)` is a place, not a publisher.
The model is not ground truth and the comparison is not a score: on several
sampled disagreements it named the journal where the article was wanted, or
the collected works instead of the essay.
Cuts a reference list into entries — at the note numbers where there are any,
at the layout where there are not — and reads each as far as its punctuation
allows. Fields that would not parse are named in `unsettled` rather than
guessed at, and `--prompt` asks a model about exactly those and no others.
Over 202 real reference sections from the corpus: 8,566 references, 95% with
a title, 82% with a year, 46% settling every field with no model involved.
A title is refused where the entry was never divided, so no record reports
its own text back as the work it cites.
`ibid.` is flagged and left alone — it points at a work, and the entry does
not say which. A rule of dashes is filled in, because the style defines it as
the author above.
chunkmd gains `note_starts` and `cites_something`: recognising a page of
notes and splitting one are the same knowledge, and two copies of that
pattern would drift.
funkpower