Reference

The corpus

Nine items on a single theme — the Los Angeles stage and its picture palaces, c. 1910–1945 — chosen to span the failure modes that matter in archive digitisation: multi-column newsprint, ornamental display type, clean printed captions, photographic postcards, a multi-page PDF, and archival TIFF.

Every item is in the public domain in the United States. The rights reasoning for each is given below, so this is a record of why each may be redistributed, not just where it came from. None carries an attribution requirement, but crediting the holding institution is good practice and costs nothing.

Click any item to see it full size — the difficulty claimed for each is much easier to judge against the actual page than against a description of it.

Inside Facts of Stage and Screen — front page 31 May 1930 · inside-facts-1930-cover.jpg
Heavy display masthead over dense multi-column newsprint, on aged low-contrast paper. The hardest common case in archive digitisation.
Inside Facts — page six 31 May 1930 · inside-facts-1930-page-six.jpg
Multi-column body text broken up by ruled advertisement boxes. Tests reading order without the masthead complicating it.
The great Victorina Troupe — vaudeville poster c. 1914 · kar-mi-troupe-poster.jpg
Curved and arched hand-lettered display type, colour on colour, tiny caption text. Nothing local reads this.
Carthay Circle Theatre premiere — postcard front c. 1930–45 · carthay-circle-premiere.jpg
Caption reversed out of a dark sky over a textured linen ground.
Carthay Circle — postcard reverse c. 1930–45 · carthay-circle-postcard-back.png
Text rotated 90°, a faded rubber stamp, handwriting and large empty areas. Clean printed type, awkwardly arranged — where free tools do well.
Grauman’s Chinese Theatre — postcard c. 1930–45 · graumans-chinese-theatre.jpg
Colour halftone with marquee lettering right at the edge of legibility.
The King of Kings — souvenir programme 1927 · king-of-kings-souvenir-1927.pdf
21-page PDF: tinted grounds, script headings, drop caps. The only multi-page item, and the only one EasyOCR and PaddleOCR cannot read at all.
Hollywood Music Box — Ken-Geki playbill 18 June 1926 · hollywood-music-box-playbill-1926.tif
Dense small type over a full sheet: cast lists, plot synopses and ticket prices at footnote size. Photographed on a dark mount, as museum holdings usually are.
Kinema Theater — newspaper display ad 2 October 1920 · kinema-theater-ad-1920.tif
Hand-lettered headlines, a programme column in tiny type, and body text flowing around an illustration. Tesseract reads 95 words of the column and misses the masthead entirely.

The thumbnails above are generated from the corpus by tools/generate_corpus_thumbnails.py and committed, so the page renders without a build-time dependency on poppler. They are not what the backends see — OCR always runs against the full-resolution originals in evaluation/corpus/images/.

Provenance and rights

File	Source	Rights
`inside-facts-1930-cover.jpg`	Internet Archive / Media History Digital Library	US work published 1930; copyright expired 1 January 2026
`inside-facts-1930-page-six.jpg`	as above	as above
`kar-mi-troupe-poster.jpg`	Wikimedia Commons, from the Library of Congress theatrical poster collection	Library of Congress: no known restrictions on publication; published pre-1929
`carthay-circle-premiere.jpg`	Digital Commonwealth / Boston Public Library	Published without copyright notice; BPL states “No known copyright restrictions”
`carthay-circle-postcard-back.png`	as above	as above
`graumans-chinese-theatre.jpg`	Digital Commonwealth / Boston Public Library	as above
`king-of-kings-souvenir-1927.pdf`	Internet Archive	US work published 1927; copyright expired 1 January 2023
`hollywood-music-box-playbill-1926.tif`	Workman and Temple Family Homestead Museum Collection, via USC Digital Library	US work published 1926; public domain on date alone. The museum’s own note is weaker — see `SOURCES.md`
`kinema-theater-ad-1920.tif`	Los Angeles Evening Express, 2 October 1920, via a newspaper-archive export	US newspaper published 1920; public domain. The scan’s producer is unidentified — see `SOURCES.md`

The full item titles and issue details are in evaluation/corpus/SOURCES.md.

Modifications

kar-mi-troupe-poster.jpg was downscaled from the source 4104 × 5116 to 1123 × 1400 so that it sits in the same resolution band as the rest of the set. Resolution materially changes OCR scores, so fixtures are kept comparable. No cropping was applied — the Library of Congress colour calibration bar and ruler remain in frame, which is representative of real archive scans.
carthay-circle-postcard-back.png was converted from the source JPEG so the set exercises the PNG path. No resampling.
The postcards were fetched at 1400 px wide via the Digital Commonwealth IIIF endpoint. Everything else is unmodified from the source.

Notes

1930 is a recent public-domain entry. Inside Facts fell into the public domain on 1 January 2026 under the US 95-year term. That is the least-aged item in the set by a wide margin.
TIFF is now covered, and testing it found a real gap. Both TIFFs read correctly on every backend, so the advertised support was honest for single-page files. It was not for multi-page ones: Pillow opens a multi-frame TIFF at frame 0, so a scanned pamphlet returned its cover as though it were the whole document. Tesseract now reads every page, as it already did for PDFs; the backends that cannot refuse by name rather than truncating silently. See the engines.
Neither TIFF is multi-page. Both are single-frame, so the corpus proves the format works and does not exercise the paging path. That is covered by generated fixtures in tests/test_tiff.py instead. A real multi-page archival TIFF would be a worthwhile addition.

Refetching

Internet Archive, Wikimedia Commons and Digital Commonwealth are all scriptable — stable download URLs, JSON APIs, no bot protection. Digital Commonwealth images come from https://iiif.digitalcommonwealth.org/iiif/2/commonwealth:<id>/full/1400,/0/default.jpg.

loc.gov and Chronicling America sit behind Cloudflare bot protection and return HTTP 403 to curl even with a browser user-agent, so anything sourced there has to be downloaded by hand through a browser. This is why the Library of Congress poster is taken from the Wikimedia Commons mirror rather than directly, and why no Chronicling America newspaper page is included despite the Los Angeles Herald being an obvious thematic fit.