Separate inference from the surrounding workflow
Scribeski is an open-source macOS app for turning a remote social-work session into a reviewed EHR draft. Its 0.1.0 release uses on-device transcription and a local note model. The published evidence is from synthetic pilot tests.
Local processing describes the transcription and inference path. The remote call still uses its call provider, and the EHR page can receive the fields filled into it. A page may transmit field changes before a final submit action.
The defensible boundary is that Scribeski does not need a separate hosted model to transcribe or draft the note. An agency still needs to assess its conferencing service, EHR behavior, permissions, and retention configuration.
Capture tracks before asking a model to interpret them
The project describes separate capture of the worker's microphone and the call application's audio. Track origin provides the initial speaker labels. That is different from assigning speakers solely through diarization of a mixed recording.
The pipeline uses Parakeet for local speech recognition, with an Apple speech-engine fallback.
Track separation and echo handling are useful controls, but accuracy on real calls must be measured with overlap, interruptions, and noisy microphones.
The session is bound to a selected client chart. Treat that binding as an input to every later write, not as an incidental label in the transcript.
Extract fields with evidence and explicit unknowns
The local note model maps transcript material into the form's fields and associates proposed answers with supporting quotations. Fields requiring judgment or lacking evidence should remain for the reviewer.
Evidence references make a draft inspectable; they do not prove that the quotation supports the chosen field or value. A pilot should score the field answer, its source quote, and whether the system should have abstained.
The extraction evaluation dated September 22 and updated September 23 reports two held-out synthetic transcripts with 106 correct answers, five unsafe answers, and 14 safe blanks. Those are counts under the repository's scoring definitions.
The same report records 182-second and 391-second held-out runs and an unmet speed gate on longer tuning sessions. Two synthetic transcripts cannot establish field performance across agencies, accents, or real sessions.
Treat the EHR write as a separate security boundary
Scribeski fills fields in Safari and leaves submission to the worker. The documented design checks client identity before writing and provides review and undo controls. Agencies should test those behaviors against their own EHR.
The security review states that the page script shares the EHR's JavaScript environment. The application therefore trusts that page. Not pressing submit does not prevent the page from observing a filled value or implementing autosave.
Use a test chart to exercise wrong-client detection, navigation during filling, validation errors, partial writes, undo, and autosave. Confirm which fields reach the server at each step before using real client data.
Inspect the configured retention behavior
The project documents no audio recording by default, with optional encrypted retention modes. Its security design also covers encrypted transcripts, session keys, and exclusions from ordinary indexing and backup paths.
Those controls have scope limits. Key deletion does not establish that every external copy is gone, and a synthetic disk-write test does not prove the absence of every possible persistence path on every Mac.
Review the project's SECURITY.md and agency deployment instructions. Record the selected audio policy, transcript lifetime, device encryption, update process, and support procedure. Test deletion and recovery with synthetic sessions.
Run a pilot that can reject the system
Start with consented role-play, synthetic identities, and an isolated EHR test environment. Have reviewers prepare reference transcripts and field answers before looking at the generated drafts.
Measure transcription errors, speaker attribution, supported field accuracy, unsupported answers, blanks, extraction latency, and total correction time. Keep a held-out set separate from prompt and form-mapping adjustments.
Define stop conditions before the pilot, including wrong-chart writes and unsupported consequential fields. Progress to a broader trial only when the measured results and the agency's review support it.
The repository lists Apple silicon, macOS 26 or later, and 32GB of recommended memory for the default note model. Check the release instructions for the exact build being evaluated; this article describes the 0.1.0 pilot evidence.
Frequently asked questions
Does local inference mean session data never reaches a network?
No. The call platform and web EHR are separate boundaries. Scribeski performs transcription and note generation locally, but filling an EHR can expose values to that page and its server.
Do the published tests establish clinical readiness?
No. The cited evidence is a small synthetic pilot evaluation with remaining quality and speed issues. An agency-specific assessment and human review are still necessary.