Skip to content

Three last things the corpus was still catching - #41

Merged
jepegit merged 1 commit into
mainfrom
last-five
Sep 9, 2026
Merged

jepegit merged 1 commit into
mainfrom
last-five

Conversation

@jepegit

@jepegit jepegit commented Sep 9, 2026

Copy link
Copy Markdown
Owner

A marker after more than one closing bracket

Parentheticals nest:

5 wt% of Super P carbon (Tim-cal) in N-Methylpyrrolidone (Aldrich)).19 Electrochemical tests…

Two brackets close at once between the sentence's last letter and its full stop, and matching a single one was still one short. They are consumed and put back now rather than looked behind — a lookbehind in Python has to be a fixed width, and a run of them is not.

A designation that begins where a number ends

A hydrate is written with the ether welded to the stoichiometry — AlH3·0.25Et2O — and the designation pattern refused to start on a token with a digit in front of it. So the plain pass said "zero point two five" and left the ether spelled out as a number.

The leftmost match still wins, so a coefficient in front is claimed with its token as before: 4AlH3 → "four AlH three". Tested.

The Norwegian company suffix is not the English word "as"

a/?s, written under re.I alongside university and institute, matched as. Every front-matter block containing that word was called an affiliation — including one paper's title:

"Aluminum hydride as a hydrogen and energy storage material: Past, present and future"

Triage dropped the title as content-free, the orientation went on saying it (which is what STR-01 asks for), and COH-01 reported the result as a leak.

While there: two hints either side of it had never matched anything at all. The pattern ends in a word boundary, and there is none between Dept. and the space after it, nor after Inc. at the end of a phrase.

Corpus

run 1 before after
total errors 329 5 2
clean papers 1/12 8/12 10/12
crashes 0 0 0

The two that remain are both understood and both narrow: a DOI fragment inside a reference line, and a second, much poorer scan of a paper the corpus already contains, where only one superscript marker survives extraction — too few for MIN_SUPERSCRIPT_EVIDENCE to let the stripper touch anything, which is the gate working as designed.

🤖 Generated with Claude Code

A marker after more than one closing bracket. Parentheticals nest -- "5 wt% of
Super P carbon (Tim-cal) in N-Methylpyrrolidone (Aldrich)).19" closes two of them
between the sentence's last letter and its full stop -- and matching a single one
was still one short. The brackets are consumed and put back now rather than looked
behind, because a lookbehind in Python has to be a fixed width and a run of them
is not.

A designation that begins where a number ends. A hydrate is written with the ether
welded to the stoichiometry, "AlH3·0.25Et2O", and the pattern refused to start on
a token with a digit in front of it -- so the plain pass said "zero point two
five" and left the ether spelled out as a number. The leftmost match still wins,
so a coefficient in front is claimed with its token as before.

And the Norwegian company suffix, which is not the English word "as". Written
a/?s under re.I alongside "university" and "institute", it matched every
front-matter block containing that word -- including one paper's title, "Aluminum
hydride as a hydrogen and energy storage material". Triage dropped the title as
content-free while the orientation went on saying it, which is what rule STR-01
asks for and what COH-01 then reported as a leak.

Two hints either side of it had never matched anything at all: the pattern ends in
a word boundary and there is none between "Dept." and the space after it, nor
after "Inc." at the end of a phrase.

Corpus: two errors across the twelve papers, and ten of them build clean.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
@jepegit
jepegit merged commit c391092 into main Sep 9, 2026
4 checks passed
@jepegit
jepegit deleted the last-five branch September 9, 2026 01:39
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant