Local inference is only one stage of a document tool. A product claiming local document processing should account for import, extraction, recognition, indexing, generation, storage, diagnostics, and optional services.
Map the actual data path before writing the privacy explanation. Users should not need to infer whether a scan or extracted passage leaves the device because the model itself runs locally.
Inventory each transformation and destination
Record source files, page images, recognised text, metadata, embeddings if used, prompts, outputs, logs, crash reports, and exports. Identify which component processes each item and whether any network destination receives it.
For macOS source processing, Apple's PDFDocument and Vision text-recognition documentation provide primary references. Verify the application's chosen path rather than generalising from one API's behaviour to the whole product.
Trace a fictional private report
Use a harmless document containing distinctive fixture text. Import it, extract or recognise it, generate a draft, save the result, trigger a controlled error, and export. Inspect the actual network and diagnostic behaviour using authorised tools and the relevant implementation evidence.
Test disconnected operation, missing model, update check, optional support export, and any cloud choice. A local happy path does not prove an error handler avoids sending the full document to a remote reporting service.
No privacy audit result is claimed here. The exercise defines evidence needed to support a bounded statement about the implemented data flow.
Separate local processing from retention and access
Data remaining on a laptop still needs account, file, backup, and cleanup rules appropriate to the setting. Generated outputs and temporary page images can contain the same private information as the original file.
Explain optional transfers before they occur. If support needs diagnostic material, minimise it and use an authorised controlled route rather than automatically attaching the entire workspace.
The tradeoff is useful diagnostics and optional capability versus a simple local-only promise. A product can choose either boundary, but the implementation and explanation must agree. Avoid claiming no data leaves the device when updates or telemetry transfer relevant content or identifiers beyond that claim's scope.
The ChalkLens portfolio account documents local source processing and inference. This private-report trace is proposed guidance and does not claim an independent audit of that application.
Check exported files and temporary artefacts under the same access policy.
Questions about local document processing
Does a local model prove the OCR is local?
No. Inspect the actual extraction and recognition implementation and its network dependencies separately.
Is local storage automatically private?
Not by itself. Access, device sharing, backups, exports, and retention still require appropriate design and policy.
What should a review inspect first?
Trace one controlled document through success, failure, diagnostics, and optional services. Bring mismatches to a document data-flow review.