Responsible use

Untrusted by default

• LLM outputs should be treated as suggestions, not facts.
• Downstream systems must validate and constrain what is executed or shown.
• Policy packs and validators encode fail‑safes around sensitive domains.
• Every decision is logged with enough context to reconstruct what happened.

Human‑in‑the‑loop

• High‑risk actions (clinical, financial, legal) should include explicit review.
• LSAS defines explicit HITL escalation decisions instead of silent auto‑approval.
• Reviewers can use LSAS risk context, rule hits, and suggested remediations alongside their own processes.
• The schema includes cases, approvals, and attestations so overrides and approvals can become part of your audit trail.

Out‑of‑scope decisions

Even with LSAS in place, some decisions should not be delegated to a model or to an automated workflow built on top of a model.

• Final clinical diagnosis or prescribing decisions.
• Final underwriting, credit, or pricing decisions for individuals.
• Employment decisions such as hiring, firing, or promotion.
• Legal advice presented as authoritative or binding.

Logging, review, and red flags

• Every high‑risk workflow should have a clear owner and review cadence.
• Watch for patterns like repeated overrides, escalating risk scores, or blind spots.
• Integrate LSAS events with your incident response and ethics review processes.
• Pause or roll back a policy pack if telemetry shows unexpected behavior.

Safe testing guidelines

• Use synthetic or de‑identified data when exploring new patterns.
• Run adversarial tests inside your own boundary, not against third parties.
• Start in sandbox environments with more permissive policy packs, then harden.
• Use stored telemetry and sandbox environments to understand the impact of new validators and policies before rollout.

Deployment best practices

• Integrate LSAS into your existing change‑management and CAB processes.
• Define clear roles for security, compliance, product, and operations.
• Document which policy packs are in use for each app and environment.
• Align this page with your internal responsible AI or model governance standard, and reference it in runbooks.