Using gemini-3-flash-preview in batch processing mode, I started getting a very high false-positive rate for RECITATION (plagiarism) when running OCR of public domain books that are hundreds of years old. Previously I had 1/2000 false positives, I now have 1/50, which makes the results unusable despite the now much higher price of processing (2.5x more than previous books).
The prompt was changed to specifically indicate this is a Public Domain book and even includes the book title, but this does not appear to make any difference. Of course I cannot change to a higher temperature - the entire point is fidelity. I tried rerunning with the failed pages, and 90% still came back as false positives. I cannot run this ten times to finally get good results.