Archive OCR project estimator

How it is calculated

• Page size by capture class: 0.35 MB for clean 300 dpi text, 0.9 MB for mixed text and diagrams, 1.6 MB for photographed pages, 12 MB for large-format logs and maps.
• Base OCR accuracy by the same class, reduced by the handwritten share and by each language past the first.
• Pages falling below your confidence threshold go to the review queue; the queue size drives the reviewer-months figure.
• Derived artefacts — extracted text, layout JSON, a searchable PDF layer — are taken as 35 % on top of the images.

Where this stops being true

Accuracy figures are representative, not predictive. A single archive can contain crisp typed reports, faded thermal prints, blueprints, and a driller's handwriting in two scripts, and the average is meaningless across that mix — the only way to know is to run two hundred representative pages through the actual pipeline before committing to a schedule. The review-queue model here is also deliberately crude: real queues grow when reviewers disagree with each other, which no estimate of confidence captures.