Source-Locked Content Pipeline

Document pipelines often emit output that looks more complete than the source permits. OCR can truncate a word, merge columns, or attach a rationale to the wrong item. A language model can make the error look coherent.

The work is to keep those layers visible: what the source contains, what an engine returned, what a human corrected, what remains ambiguous, and what an independent pass concludes. Import fidelity is not medical correctness. A package can faithfully represent a source that is outdated or wrong.

Two surfaces, one system. Ingestion preserves document identity, page boundaries, chunk provenance, transformation history, and replayability. Assessment packaging preserves item order, wording, choices, printed answers, rationales, images, and structure — and refuses to invent missing material.

When engines disagree or a companion item is not in the visible source, the record is quarantined instead of completed. Quarantine is a product feature. A system that can preserve unresolved uncertainty is safer than one that can fill every field.

Access is local. Public examples are synthetic. Proprietary stems, keys, scans, and recognizable assessment screenshots are not published.