New from the Lab·The Compass — an open moral reasoning standard for AI, tested across frontier modelsExplore →
Production AI Institute · PSF v1.1 open standard
AI Right-To-KnowAI Data Use IndexCheck My AI ToolsPolicy Change WatchAgent ReadinessPublic BenchmarkContact
Deployment Safety Assessment

Apply for the PAI Production Safe Mark

An independent assessment of your AI systems against the PAI Production Safety Framework. Transparent criteria, documented review, public verification record.

Standard assessment timeline (15 business days)

Day 1–2
Application reviewed, scoping call scheduled
Day 3
Scoping call — agree scope, timeline, evidence
Day 4–7
Self-service workflow mapping using PAI Workflow Studio
Day 8
Evidence submission deadline
Day 9–13
Assessor review period (5 business days)
Day 14–15
Report delivered. Mark issued on pass.

Pass threshold

Each domain is scored 0–5. To pass, you need at least the threshold shown for every PSF domain. No domain may score 0.

PSF-1Input Governance≥3.0/5.0
Prompt injection controls, input validation, rate limiting
PSF-2Output Validation≥3.0/5.0
Schema checks, hallucination controls, safe fallbacks
PSF-3Data Protection≥3.0/5.0
PII handling, lawful basis, retention and deletion
PSF-4Observability≥3.0/5.0
Logging, quality monitoring, drift and cost alerts
PSF-5Deployment Safety≥3.0/5.0
Change control, canaries, rollback and incident runbooks
PSF-6Human Oversight≥3.0/5.0
Autonomy limits, review queues, override mechanisms
PSF-7Security≥2.0/5.0
Least privilege, key management, tenant isolation
PSF-8Vendor Resilience≥2.0/5.0
Fallback providers, deprecation planning, SLA review

What evidence to prepare

Build the evidence room before the scoping call so the assessment feels controlled, not stressful.

Build evidence room

Technical evidence

  • ✓PAI Workflow Studio workflow maps of in-scope AI systems
  • ✓Monitoring dashboard screenshots (redact PII)
  • ✓Input validation / output safety code or config
  • ✓Access control and API key management docs
  • ✓Data retention and deletion capability evidence

Process evidence

  • ✓Incident response runbooks (AI-specific)
  • ✓Post-incident review examples (if any)
  • ✓Human oversight / review queue documentation
  • ✓Vendor SLA review and fallback planning
  • ✓Consent and data handling policy excerpts
No payment until assessment is scheduled and scope is agreed.