@drvibeai/phi-guard

A tiny, dependency-free toolkit that lets an agent manage PHI posture instead of just disclaiming it. Three functions:

• complianceGate({ deployment }) — decide the operating mode from the host AI's BAA status: full (BAA-covered) or minimized (not / unknown), with guidance and a route to compliance review.
• redact({ record, text }) — tokenize direct identifiers (name, email, phone, DOB, address, MRN, SSN) locally; scrub emails/phones from free text. Returns the redacted data plus a map (token → real value) you keep on the user's machine.
• rehydrate({ text, map }) — restore the real identifiers into the approved draft, locally, right before a human sends it.

The model reasons over [NAME-1]; the real value never leaves the device.

Honest about what this is. This minimizes direct identifiers (best-effort). It is not certified HIPAA Safe-Harbor de-identification — free text can hide identifiers and quasi-identifiers (city/ZIP, rare conditions) may remain. The compliant path for real PHI is a BAA-covered deployment. A human must review redacted output. See the repo DISCLAIMER.

Install

npm install @drvibeai/phi-guard

Use

import { complianceGate, redact, rehydrate } from "@drvibeai/phi-guard";

// 1) Gate first.
const gate = complianceGate({ deployment: { baa_covered: false, vendor: "ChatGPT (free)" } });
// gate.mode === "minimized"  -> redact before the model sees anything

// 2) Redact locally; keep the map on-device.
const { redacted, map } = redact({
  record: { full_name: "Jane Doe", contact_email: "[email protected]", reason_for_visit: "follow up" },
});
// redacted.full_name === "[NAME-1]"   (map["[NAME-1]"] === "Jane Doe")

// 3) The model drafts with tokens; a human approves; rehydrate locally before sending.
const sent = rehydrate({ text: "Hi [NAME-1], reach you at [EMAIL-1]?", map });
// sent.text === "Hi Jane Doe, reach you at [email protected]?"

Why

Patient-facing agent work is inherently identified — you can't "de-identify and still follow up." phi-guard lets the reasoning happen over de-identified text while identifiers stay local, and routes naive users toward a BAA-covered deployment when they need real PHI. It's the reusable pattern behind DrVibe's patient-facing skills.

Test

npm test   # offline, deterministic

Apache-2.0 © DrVibe.ai