ocr

Jeffrey Emanuel (@doodlestein) — 3h It's so annoying when you get "content filtering" API errors from Anthropic for totally harmless stuff. I'm working on my FrankenOCR project now, and as part of that, I'm using a PDF of an old book (Clowes' History of the Royal Navy) because it's public domain and very challenging for OCR, with lots of tables, weird typesetting, etc. I needed to get some ground truth content to test the output of my tool, so I asked Claude to use its internal vision ability to do that. Everything was working fine until it got to a page that apparently offended its sensibilities. I speculated that it might be because it's about war, but Claude responded: "Good instinct, but I don't think it's "war" broadly — page 107 (voyages/exploration) transcribed fine. It's most likely either the large verbatim transcription in my visible prose, or the specific page 108 narrative (mutineers casting sick men adrift to die). The fix is simple: I'll keep book text out of my prose and put ground truth only in tool calls (the page-107 Write succeeded), and pick neutral pages. Let me keep going without quoting the book." I'll keep it in mind next time I'm planning to cast sick men adrift to die that I shouldn't run my plans by Claude for advice. [Embedded screenshot: split view showing a Claude conversation transcript on the left (partially legible: "...crate ground truth (de-hyphenated reading text) for both. Page 107 (book p.84): ...we made a careful examination of the western coach of Spitsbergen, and on his way home discovered the island close under the ice, having sent the vessel southward...") and a scanned title page from Clowes' "History of the Royal Navy" (Wm. Laird Clowes, Vol. II) on the right.]