[daveguy]

> which is supposed to be somewhat universal in correcting similar OCR'd PDFs

Xerox would like a word.

https://news.ycombinator.com/item?id=29223815

Point being, "correcting" to "correct looking" may be worse than just accepting errors. Errors are often clearly identified by humans as a nonsense word. "Correcting" OCR can result in plausible, but wrong results that are more difficult for the human in the loop to identify.

[dperfect]

That's true if we're correcting OCR of actual output text. In this case, it's operating on the base 64 text, trying to produce chunks that form valid zlib streams and PDF syntax so the file can be intact enough to be opened. "Just accepting errors" would mean not seeing any content in the file because it cannot be read.

So yes, the "fixed" output has errors, but it’s not hallucinating details like an LLM, nor is it trying to produce output that conforms to any linguistic or stylistic heuristics.

The phrase "correcting similar OCR'd PDFs" should have been "correcting similar OCR'd base 64 representations of PDFs".