The mixed-script OCR rule needs care at a finer level than the page. A Hindi circular can have English course codes and numeric fee columns, so selecting one full-page OCR result because its confidence is higher can discard the very cells a user asks about.

I would compare the alternatives per line or table cell, retain uncertain tokens for review, and test code/amount pairs rather than average page confidence alone. Your saved bounding boxes make that review practical: a disagreement can open the exact cell in the scan, without asking a reviewer to inspect an entire circular.