Skip to main content

Technical Overview

Clinical Corvus was built to be useful in real acute care conditions: incomplete interoperability, automation bias risk, and strict privacy requirements.

Principles​

  • Bounded support: the system knows what it cannot answer reliably
  • Limited autonomy: preferring reviewable drafts to authoritative answers
  • PHI in backend by default: identifiable context does not leave without explicit configuration
  • Determinism where it matters: keeping clinical state consistent and auditable

Current product boundary​

The system operates within a limited beta contract:

  • Supported: clinical question with context, differential synthesis, plan drafts under clinician review
  • Warned: questions with insufficient evidence — uncertainty declared
  • Blocked: autonomous diagnosis, automated treatment planning, unrestricted advice

Clinician review is part of the normal flow, not an exception.

Architecture highlights​

Multiple execution paths​

The current dashboard exposes two AI paths that share product context but are not the same execution surface:

  • Canonical task pipeline (/api/agents/tasks/*): query, discussion, research
  • Clipboard-plan BAML pipeline: synthesis, conflict detection, plan suggestion

Public description should not imply that every AI flow traverses a single pipeline.

Shared clinical query profile​

Each request enters through a task-oriented router that constructs a reusable clinical query profile. This profile captures dimensions like task_type, decision_horizon, context_sufficiency, and answer_shape. The profile is reused across Coordinator, CDA, CRA, and Compass to avoid routing drift.

CaseState vs. longitudinal context​

The system separates:

  • CaseState: active episode reasoning state, patch-based and versioned
  • Patient Context Manager: longitudinal aggregated patient context
  • Learned user memory: clinician preferences and long-term summaries

Episode reasoning does not persist as generic learned memory.

PHI egress gate​

The backend implements a PHI egress gate that classifies content, blocks external calls when PHI is present under disallowed conditions, and records structured audit events. External research is disabled by default.

What you find here​

Conceptual and operational view of the platform: