SNEWPapers vs American Stories: 18 issues, one per decade

Eighteen newspaper issues, one per decade from the 1770s to the 1950s (American Stories has nothing for the 1780s), compared component by component. Each report has a scorecard, then every snewpapers component with the American Stories boxes inside it: crops of both from the page scan, both texts, and the verdicts of an LLM shown only the texts. Each overlay draws both segmentations on the page scans, American Stories on top and snewpapers below, with the verdicts marked.

Judged wrong: the American Stories articles and snewpapers components the LLM's decisions found wrongly segmented. Real words: the share of each side's words (3 or more letters) found in an English word list, over the same text.

All the scorecard numbers

One row per benchmark issue. snewpapers components: every match, 1 ours : N AS and ours-only case, by its verdict (see each report's scorecard). N ours : 1 AS and attached-outside questions: which side the model found right in both orders; "other" is neither, unstable or no verdict. AS only: left-over AS text the model called advertisements, terms or page furniture (correctly not served) or not (missed). Judged wrong: what those decisions judged wrong, counted in AS articles and snewpapers components (an as_split of one component into 15 AS articles decided "one" is one decision and 15 AS articles). AS errors: what the decisions for snewpapers found American Stories got wrong, with the decisions behind them: an N ours : 1 AS is N items AS merged (seven items of ours in one AS article: 7), a 1 ours : N AS N fragment articles. Text: the median fuzz ratio over 1:1 matches (the totals row: over every match). Real words: the share of each side's words (3 or more letters) found in an English word list, a long s read as f allowed, over the components of ours AS has text inside, so both sides count the same text. Pages: how many American Stories scans are the LoC scan snewpapers' boxes are on.