I tried to leave a comment, but it doesn’t seem to be showing up there.

I’ll just leave it here:

too tired to look into this, one suggestion though - since the hangup seems to be comparing an L and a 1, maybe you need to get into per-pixel measurements. This might be necessary if the effectiveness of ML or OCR models isn’t at least 99.5% for a document containing thousands of ambiguous L’s. Any inaccuracies from an ML or OCR model will leave you guessing 2^N candidates which becomes infeasible quickly. Maybe reverse engineering the font rendering by creating an exact replica of the source image? I trust some talented hacker will nail this in no time.