Security and data boundaries
LSAS Platform connects authenticated application identity, policy evaluation, bounded execution, and decision records. These controls have distinct responsibilities: a validator finding, an execution grant, and a human release approval are separate artifacts.
This page describes application controls and their limits. Infrastructure configuration and independent release evidence must be evaluated for the deployment in use.
Identity and authority
- Console services enforce current membership and role checks. Gateway requests require an active, unexpired application key; client-supplied tenant identifiers do not grant access.
- Alignment activation requires the applicable approval process. Version and configuration changes invalidate previously bound authority.
- Invitations use signed, expiring, single-use tokens. Administrative account access currently uses application authentication; customer IdP and MFA integration require separate implementation and qualification.
Bounded runtime execution
- Registered skills are bound to server-derived scope, explicit destinations, supported data shapes, deadlines, and budgets. Untrusted input cannot install skills or broaden execution authority.
- Ingress checks precede provider effects. Outbound transport constrains destinations, redirects, response sizes, retries, and cancellation.
- Blocked, abstained, and held results do not release protected output through the governed gateway path. Unsupported clinical and generated-code profiles fail explicitly.
Protected output review
- The opt-in nonclinical text profile binds an encrypted held artifact to its exact digest, application, original key, configuration, and expiry.
- Separately assigned reviewers inspect the exact artifact before approving or rejecting it. The original application consumes an approved result through the release boundary; a console approval is not itself delivery.
- The current grant lifetime is 60 seconds, including generation. Expired or changed authority requires a new execution. This bounded profile is not a clinical sign-off workflow.
Data handling
- Standard model decision telemetry records derived outcomes and findings. Other surfaces can retain submitted cases, evidence, workflow artifacts, and raw or transformed connector data.
- Provider credentials and held review artifacts use scoped application encryption. Other stored content and logs require the configured database, infrastructure, and retention controls.
- Configured external model providers and connectors receive the payload permitted for that request. Deploying LSAS Platform privately does not by itself keep every downstream request inside that network.
Operations and evidence
- Assessment jobs use durable leases, fencing, cancellation, and bounded retries. A separately operated worker executes queued assessments.
- Retention cleanup and rollups require the operations worker. Legal holds and configured retention scope apply to the implemented cleanup paths; backups and external logs have separate lifecycles.
- Structured events and an export outbox provide integration points. SIEM delivery, incident response, capacity, and recovery must be demonstrated in the target environment.
Sandbox and intended use
- Guided examples deliberately distinguish baseline observations, simulated actions, and governed behavior. Showing raw synthetic data in a diagnostic comparison is intentional.
- The real Epic sandbox path is limited to patient read/search using dedicated sandbox credentials. Simulated fixtures do not invoke Epic or a model provider.
- Pattern checks do not establish exhaustive PHI detection, factual truth, regulatory conformity, or clinical correctness. The published capability status identifies implemented and unsupported scope.