Skip to main content

Validation and beta status

Clinical Corvus is in limited beta. Current validation primarily covers software contracts, traceability, policies, and internal evaluations. It does not demonstrate prospective clinical effectiveness, comparative superiority, or regulatory authorization.

What has been verified

AreaAvailable evidence
Runtime identityThe health check can expose version and commit to detect a stale service.
Canonical pipelineGoverned tasks use /api/agents/tasks/* and produce the shared response contract.
Response railsCanonical responses can carry ClinicalEvidencePacket, PatientContextManifest, and CorvusProvenance.
EvidenceInternal evaluations inspect citations, admitted or rejected evidence, and claim-to-source binding.
TenantMigrations and services apply tenant scoping; covered PostgreSQL tables receive RLS policies.
External sharingThird-party PHI sharing is disabled by default, and public queries receive specific checks.

These items confirm mechanisms and contracts. They do not confirm uniform coverage across every application path.

Open gates

GateWhat remains to be shown
Clinical utilityComplete responses need systematic specialist review.
Source supportClaim coverage must remain consistent across questions and execution paths.
Longitudinal handoffContinuity between shifts and episodes needs operational validation.
Product surfacesMonitor, Plan, Workspace, and Clipboard must preserve context, uncertainty, and traceability end to end.
OperationsLatency, availability, support, and incident response must be measured in a pilot.
Clinical effectivenessNo prospective improvement in outcomes or care time has been demonstrated.

Interpreting a response

The interface may show states equivalent to ready for review, partial, insufficient evidence, blocked, or mandatory review. These states guide the user’s next step. They do not validate the clinical content.

Next validation stage

The next stage should combine complete clinical review packets, replay of the screens clinicians use, controlled pilots with predefined measures, and an operational audit of tenant controls, data sharing, billing, logs, and incidents.

Results and dates should be published when a stable, reviewable package exists, not when only an isolated benchmark is available.