Trust

Security

ScalpelNote is designed around de-identified clinical input, physician-controlled drafts, restricted access, and short retention for source audio and transcripts.

Last updated: 11 July 2026

De-identification boundary

  • Direct patient identifiers are prohibited and are not required by the report workflow.
  • Users confirm de-identification before every text or audio case.
  • Typed input, edits, DeskLink saves, AI output, and clean copy are checked for high-confidence direct identifiers.
  • Detected direct identifiers are blocked or replaced before clinical text is retained.
  • Transcripts are checked and redacted immediately after speech-to-text and before database storage.

Automated detection cannot guarantee complete de-identification. Audio reaches the configured speech-to-text provider before server-side transcript checking is possible, so users must never speak direct identifiers. The user must also consider whether a rare combination of otherwise indirect facts could identify a patient.

Application controls

  • Invite-only accounts and passwordless, single-use email login codes.
  • Secure, HttpOnly, host-only session cookies and server-side owner checks on reports.
  • Temporary DeskLinks whose tokens and short codes are stored only as keyed hashes.
  • Rate limits on authentication, report generation, audio upload, billing actions, and short-code redemption.
  • Administrative views designed to exclude report text, transcript text, raw audio, and AI prompt bodies.

Data protection

  • The application and its primary data stores are hosted on Hetzner infrastructure in the European Union.
  • TLS in transit and encrypted PostgreSQL backups.
  • Raw audio deletion after transcription with a 24-hour lifecycle backstop.
  • Transcript deletion after 7 days and physician-controlled report deletion.
  • Recordings, transcripts, prompts, reports, and outputs are not used by ScalpelNote for model training; production API content is not opted in to provider training-data sharing.
  • Metadata-only operational logs that must not contain clinical report or transcript text.
  • Production configuration checks that reject default secrets and unsafe authentication settings.

Clinical safety boundary

Every output remains a draft. Review markers distinguish dictated information, standard assumptions, critical assumptions, and missing required fields. Server-side gates prevent clean copy while required review items remain unresolved. These controls support review but do not replace professional judgment.

Report a concern

Send suspected vulnerabilities, accidental identifier disclosure, or security incidents to info@scalpelnote.ai. Do not include patient data in the initial message. We do not currently claim a third-party security or privacy certification and do not guarantee that any internet service or automated detector is risk-free.