Evidence records for governed AI claims
Metric labels and publication rules
Hard numbers require an evidence record and approval; otherwise product pages use the governed fallback wording.
Customer-verified
A customer outcome that cannot be promoted publicly until authorization, scope, and legal approval are recorded.
Third-party validated
An external audit, assessor, lab, or certification result. No third-party validated quantitative claims are currently published.
Internally tested
A reproducible internal test with benchmark metadata; use cautious wording unless public approval is granted.
Target architecture
Capability or architecture direction, not a measured production result.
SLO
A service objective or review commitment, distinct from historical measured performance.
Illustrative scenario
Example workload or buyer situation used for education; not a customer result.
Current proof-claim records
43% reduction in clinical errors
Label: customer-verified. Status: draft; use fallback: designed to reduce documentation and triage risk.
Benchmark metadata: date pending legal approval; version legacy case-study packet; hardware, topology, dataset size, concurrency, duration, failures, and reproduction method are not public until customer authorization and legal review are complete. Workload: customer clinical documentation workflow. P50/P95/P99 latency: not applicable.
<100ms latency
Label: internally tested. Status: draft; use fallback: low-latency orchestration for production workflows.
Benchmark metadata: date 2026-07-15; version AIOS reference architecture draft benchmark packet v0.3; hardware internal reference compute not approved for public reliance; workload synthetic agent handoff and policy-decision simulation; dataset size deterministic fixture set not approved for external publication; topology single-region target deployment with API gateway, orchestration worker, policy service, and audit log sink; concurrency 50 simulated workflow starts; duration 30 minutes; latency P50 42ms, P95 88ms, P99 126ms; failures 0 transport failures in synthetic run; reproduction method internal deterministic replay harness available only through reviewed methodology request.
Built for regulated teams
Label: target architecture. Approved non-quantitative positioning for regulated AI program needs.
Audit evidence by design
Label: SLO. Approved capability claim for structured review and audit evidence; it must not imply external certification.
Illustrative procurement evidence workflow scenario
Label: illustrative scenario. Approved as buyer education only; it is not a customer result or benchmark.
Security review response within scoped contractual SLA
Label: contractual SLA. Applies only when included in an executed agreement; public fallback wording is that security review response timing is scoped during procurement.
Third-party validated status: no third-party validated quantitative claims are currently published; any future external audit, assessor, lab, or certification result must be added here before use.