Skip to main content

Local-First Privacy & Trust

Clinical Corvus enforces a strict 'Local-First' Trust Boundary. By default, all reasoning and ingestion occur within the secure backend, minimizing the exposure of Patient Health Information (PHI).

Egress-Filtered Information Retrieval​

When external evidence is needed, Clinical Corvus uses an egress-filtered approach: external calls should carry clinical keywords, not patient-identifying context.

When the Clinical Research Agent (CRA) requires external evidence (e.g., searching PubMed or the Open Web), it employs a Local Sanitization Layer:

  1. Sanitization: A rule-based/regex layer strips PHI (names, MRNs, dates) in-process before any network request is formed.
  2. Anonymized Queries: External search providers receive only anonymized clinical keywords (e.g., "septic shock protocols", "vancomycin dosing"), never patient context.
  3. Prevention: The architecture prevents 'leakage by default' rather than relying solely on post-hoc filtering.

If an institution prefers to use its own model endpoints (e.g., a private OpenAI/Azure/other account), Clinical Corvus can be pointed to those credentials so data stays within that governance boundary.

Ephemeral Memory Architecture​

To align with data sovereignty requirements (GDPR/LGPD), the system avoids long-term retention of raw patient data on Corvus servers.

  • CaseState: Persists structured, episode-scoped reasoning state (snapshots + patches) on institutional infrastructure (local storage).
  • AgentMemoryService: Stores short-term event timelines (Redis) for conversation continuity.
  • Data Lifecycle Management: Automated TTL (Time-To-Live) sweeps sanitize ephemeral logs and age-out long-term entries, ensuring compliance without manual intervention.

Agentic Safety Layers​

Trust is not just about privacy; it's about the reliability of the clinical advice. Corvus implements active safety patterns:

1. Goal Verification​

Before returning any response, a dedicated verification loop (VerifyGoalCompletion) assesses if the drafted answer actually addresses the user's core intent.

2. The Critic Agent​

A "Producer-Critic" pattern employs a separate adversarial model to critique responses for accuracy, completeness, and clarity. It flags responses with INFO, WARNING, or ERROR severity before they reach the clinician.

3. Confidence-Based Escalation​

A formalized trust model (Low, General, High Clinical, Critical Life Safety) automatically triggers a "Human-in-the-Loop" pause (HitlPauseTool) if confidence dips below the threshold required for the specific care setting (e.g., < 0.98 for Critical Safety).

Security Posture​

  • Content-Security-Policy (CSP): Strict headers to prevent XSS.
  • Audit Logging: Deterministic logging for all PHI-touching endpoints.
  • Rate Limiting: Centralized rate limiting to prevent abuse.
  • Opt-In Compliance: For deployments using hosted models (e.g., GPT-4), third-party endpoints are reachable only if explicit Zero-Data Retention (ZDR) and BAA/DPA agreements are verified.