Extract or reconstruct canonical text.
Use native text where available. Use OCR reconstruction for image-first pages while preserving page and span provenance.
Field note · Internally tested
Prepare sensitive documents before AI sees the working copy.
HONONOH tested a local document-preparation pipeline that reconstructs scanned text, identifies sensitive spans, applies deterministic replacements, preserves a protected ledger, and blocks export when validation finds unresolved privacy risk.
Capability evidence, not a production clinical-deployment or client-outcome claim.
01Reconstruct
02Detect and review
03Replace deterministically
04Validate before export
The operating problem
Raw private records are not a safe input boundary.
Native files and image-first pages can contain identifying information that must remain traceable during preparation without leaking into downstream AI or review work.
01Scanned pages need reconstructionImage-only documents need canonical text, geometry, and provenance before text-level controls can work.
02Detection is not final authorityAutomated candidates can miss context or assign the wrong role. A person must resolve ambiguous spans.
03Replacement must remain auditableStable pseudonyms, protected ledgers, and validation evidence preserve the path from source to working copy.
The tested system pattern
Local processing with deterministic control and human authority.
Review candidates and apply exact replacements.
Automated methods support discovery. Explicit operator decisions and deterministic code remain the final replacement authority.
Block unsafe export.
Validation checks the sanitized output and protected ledger. Residual matches or leaked ledger values stop the release path.
Controlled proof run
What the current evidence demonstrates.
The proof used one native text file and one controlled image-only scanned-style PDF. It tested activation and control flow, not general document accuracy.
The pipeline produced the required review artifacts.
- OCR reconstruction report and readable reconstructed text
- span redaction ledger and measurement report
- case context, name candidates, and OCR review candidates
- final anonymization and validation reports
- passing validation with reviewed OCR decisions
The reconstruction substrate was present and usable.
- one reconstructed page
- four reconstructed usable line spans
- 95 reconstructed text characters
- five crop-evidence rows with none missing
- 3.80-second controlled activation run
Interpretation
The test demonstrates a working local control pattern for clean printed scanned text. The reconstructed Markdown and span records are the context-preserving proof substrate. Region-redacted PDFs are visual containment outputs, not the canonical downstream context product.
Evidence boundary
What this does not prove.
The label stays at Internally Tested until broader fixtures and operating evidence justify a higher state.
01No universal OCR claimColumns, tables, handwriting, signatures, low-quality scans, and difficult region splits remain open quality risks.
02No unattended approvalHuman review remains required for ambiguous identities, roles, and release decisions.
03No production outcome claimNo clinical deployment, public distribution, independent client result, or quantified operating benefit is claimed.
04No complete export-package claimThe controlled run did not produce the planned export-bundle manifest.
Have a document workload?
Start with representative cases and an explicit release boundary.
Describe the formats, volume, authoritative records, privacy constraints, accountable owner, human authority, and acceptance cases. Do not upload confidential source material.