✎ Click any word to correct it — press Enter to save, Esc to cancel.
A pipeline is an ordered list of labeled stages. Each stage's label is what defines the engine — pick a locally hosted model or an API model, and optionally override its prompt to tune it for a document type (e.g. "18th-c. Marathi Modi handwriting", "Bengali gazette, printed"). Stages run in order; later stages can consume earlier stages' output (e.g. a language-ID stage feeding a script-specific OCR stage).
Every stage's output on the standing benchmark set (ground-truth transcribed pages per script/hand-type) is scored nightly on WER/CER against HITL-verified truth. New open-weight releases are auto-queued for a benchmark run before being offered as a pipeline stage option above — see the architecture note for the full leaderboard/promotion flow.
Accuracy % (char/word) populates only after a page has a HITL-verified ground truth on file — until then only model self-reported confidence is shown, never presented as accuracy.