Why scanned PDFs need human review

Not every official PDF can be read by a machine. In our source research on 12 October 2026 we met three different situations, and each one needs a different, honest response.

Three kinds of PDF

• Text PDFs in Unicode. Bangladesh Bank's English circulars and NBR's recent SROs (generated from word processors) carry a usable text layer. Extraction works; search, diff and citation work.
• Text PDFs in a legacy encoding. Bangladesh Gazette PDFs carry text, but encoded in Bijoy/ASCII, a pre-Unicode scheme. The characters extract, but as the wrong letters. Without conversion, a search for an SRO number finds nothing and a quote is unreadable. NBR's Bangla PDFs can also need glyph-order normalisation (Nikosh) before indexing.
• Scanned image PDFs. Some BSEC directives are scanner output with no text layer at all. Of the two BSEC files we sampled, one was scanned and one was text. Nothing extracts until OCR runs, and Bangla OCR on scans of mixed quality produces errors in exactly the places that matter: numbers, dates and negations.

Why a confident summary is the wrong answer

A summary produced from garbled or OCR-guessed text looks the same as a good one. The reader cannot tell that "not later than 15 days" became "later than 15 days", or that a rate of 7.5 became 75. The cost of a wrong confident sentence is far higher than the cost of a visible "pending".

What we do instead

• Every document and version carries an extraction quality: text, partial, OCR needed, manual review, failed.
• Scanned files are shown with their official link and the label "scanned, summary pending". No brief is generated from them automatically.
• Low-confidence items go to a human review queue. A reviewer either writes the summary from the original or marks the item as needing the regulator's text version.
• Citations are only ever quoted from text we extracted with acceptable quality. If we cannot quote it, we do not assert it.

What you can do

If you receive a scanned instrument that matters, ask the issuer for the text version or the Gazette reference, and treat any machine summary, ours or anyone else's, as a pointer rather than a source.