# Explicit scanned-PDF corpus OCR

The public `source_corpus` tool can process an acquired PDF population with
`begin` using `filters: {channel: "ocr_text"}` and optional `ocr_language`
(default `eng`). Repeated `process` calls resume a durable per-source page
cursor. Installed CPU Poppler and Tesseract are required. The default channel
remains `native_text`; completing native extraction of an empty scanned page
does not complete its OCR extraction.

Each OCR page retains the original PDF identity, rendered PNG, full Tesseract
text and TSV geometry in durable content-addressed files, with exact hashes,
page lineage, engine version and language. Immutable segmented evidence powers
bounded search and read operations. OCR receipts preserve the original issuing
owner; subsequent reads bind their exact exposed range to the current owner.
OCR TSV geometry is separate from native PDF geometry.

`corpus_work` can explicitly select the complete `ocr_text` population for
anchored extraction or summary. A read checkpoint is prepared when text is
returned, then confirmed only after the production runner successfully sends
the exact protected OCR text carrier to its configured backend. Failed
transport cannot confirm that read or authorize a derived commit. Derived
content remains `actor_attested`, requiring independent factual review.

OCR completion means the engine processed every selected page. It does not
prove transcription accuracy or visual comprehension: `visualCoverage`
remains `not_assessed`, and OCR does not receive native evidence authority.
Complete OCR-derived text exports can pass the separate export reviewer after
it reopens every original PDF, rendered page, OCR text, TSV and immutable
source/producer/exposure record. Fresh bounded factual review sees explicit
`ocr_text` transcription provenance and the unassessed visual limitation.
The complete source and output unions remain required before retention and
provider delivery. Exact numeric source-line transcription can be verified
against OCR engine output; this does not verify that output against the image.

Native table-cell, count, sum and aggregate quantity gates remain strict: OCR
text and TSV do not confer native field or arithmetic authority. OCR-derived
arithmetic needs a separately reviewed transcription/operand population before
approval; matching digits or a prose verdict cannot provide it. The inline
Telegram source-answer review path still has no OCR-specific source-index
adapter; the verified delivery route here is the full derived text export.
Both OCR-only and mixed native/OCR factual inline candidates are refused after
an actual successful OCR model handoff. The barrier is durable for the admitted
session/run/epoch; ingestion, metadata and failed transport do not create it.
The refusal preserves existing OCR evidence and gives the exact work/export/
review steps instead of requesting another native-empty read.

After fresh `corpus_export_review` completes under the current request, use
`task_complete` with `deliver_artifacts: true` and the exact neutral notice
`The reviewed OCR results are attached.` The host reopens the complete export,
its original OCR source proof and current signed semantic receipt. Completion
still requires actual provider acknowledgement for every selected output.
That file review does not approve unrelated inline factual prose, and the
neutral notice alone cannot waive missing review or delivery receipts.

Population traversal and page work are resumable and have no total source or
page count ceiling. A single-page rasterization still needs physical CPU/RAM
resources; this implementation does not tile unusually large page images.
