AI Transmission Path Audit

Audit date: 2026-04-23

Scope: The /api/generate Vercel Edge Function and the full request path from the mobile client through the proxy to OpenAI. The audit covered the request body, the request headers, the proxy's logging, the analytics and telemetry that fire on the narrative-generation path, and the Vercel platform configuration.

Methodology: Static review of every file involved in constructing the prompt and every logging call on the request path, followed by dashboard verification of the Vercel project configuration and an empirical sample of production runtime logs.

The body of the request sent from the mobile app to /api/generate contains four kinds of content:

• Structured enums (request type, model name) — pre-enumerated, no user-typed content. Safe.
• Structured chip values (interventions, hallucination types, safety status, Activities of Daily Living (ADL), ambulation, eating, medication compliance, commitment status, and all clinical assessment enums) — passed through whitelist label maps before being embedded in the prompt. They cannot carry user-typed patient data. Safe.
• Numeric and date-typed fields (vitals, Morse / Braden / Glasgow Coma Scale (GCS) scores, wound measurements, pain intensities, infection screening counts, note date, note time) — structured types with no free-text. The note date and note time are the wall-clock time of note generation, not patient Date of Birth (DOB) or admission date. Safe.
• Free-text fields (per-section observation, notes, details, direct quotes, education barriers, goal progress notes, wound notes, signed nursing notes) — transmit verbatim. There is no automated client-side de-identification step before transmission. See Known Tolerances below.

The structured data model is the primary protection at the structured-field layer: there are no fields for patient name, Medical Record Number (MRN), or date of birth in the chart data model. Free-text fields transmit as the nurse enters them.

Three request headers leave the device:

• Content-Type — literal application/json.
• An anonymous RevenueCat user identifier ($RCAnonymousID:…) used for subscription validation. Pseudonymous; not patient PII.
• A free-narrative entitlement boolean (sent only when the free-pool path is used).

No patient data, free-text content, or device fingerprint travels in the headers.

Every console.* call site on the request path was reviewed. Only request metadata is logged:

• Request type (narrative / freeform / clinical) and the model name being used.
• An 8-character prefix of the anonymous customer identifier — sufficient to correlate a small number of recent requests during incident triage, not enough to reverse to an individual.
• Network-level error messages from the upstream OpenAI iterator and from the client retry path (e.g. “network request failed”).

Request and response bodies are not written to logs. The Vercel Edge Runtime does not capture request or response bodies in runtime logs by default, and our application code does not introduce any additional body-logging.

Two analytics events fire on the narrative-generation path:

• Narrative generated: carries a non-reversible 32-bit hash of the chart UUID. The chart UUID itself is client-generated and contains no patient content.
• Free-narrative used: carries an integer narrative count, the free-narrative limit, and a feature name literal.

Mixpanel and AppsFlyer providers are typed pass-throughs: they forward only the explicitly filtered properties from the call site, do not spread arbitrary context, and validate against a hardcoded event allowlist. AppsFlyer runs in anonymized mode.

No analytics event on the narrative-generation path carries the prompt content, the request body, or the generated narrative text.

In addition to the application-code review, the Vercel project configuration was checked on 2026-04-25:

• Log drains: none configured. No external service captures runtime logs from the project.
• Third-party log integrations (Datadog, Axiom, etc.): none installed.
• Runtime log retention: 1 day (Vercel Pro default), corroborated by querying the runtime-log API with a 24-hour window (succeeds) versus a 7-day window (fails).
• Project access: scoped to the development team with two-factor authentication enforced on owner accounts.

The audit included an empirical sample of production runtime logs over a 24-hour window. Across roughly 200 production log lines, full-text searches for prompt header strings and free-text section markers (the literal tokens our formatters emit when free-text is present in the body) returned no matches in /api/generate request lines. The one matching log line was the audited-safe metadata form (request type and a truncated model name only), confirming the static-code finding against live traffic.

We disclose three deliberate tolerances in the transmission posture. None of them is a surprise; all three have been considered and accepted as the cost of the current architecture.

• Every field in the request body is accounted for and categorized.
• Request headers carry no patient data.
• Proxy logging carries no request body content under current code — verified empirically against a 24-hour sample of production runtime logs.
• Analytics events on the narrative-generation path carry no prompt or narrative content.
• No Vercel log drains and no third-party log integrations are configured for the project.
• Free-text fields transit to OpenAI verbatim; Nurse Charting Pro does not represent the OpenAI payload as HIPAA-de-identified under 45 CFR §164.514. Facility deployments rely on the upstream Business Associate Agreement (BAA) + Zero Data Retention (ZDR) chain.

As a result of this audit we corrected several pieces of internal documentation and sales copy that previously described an automated client-side de-identification step that did not match what the code does, and we updated the public web copy on the privacy, security, HIPAA, and FAQ pages to disclose the request-scoped AI transmission path rather than relying on a blanket “PHI never leaves the device” framing.