Evidence taxonomy

Every retained result should name its command, observed output, source SHA, environment, and profile. A label such as “integration” is not enough; the exercised boundary determines the claim.

Evidence class	What it proves	What it does not prove
Static/type/lint	Selected source parses, types, lints, or satisfies the named static rule	Execution or a completed workflow
Focused unit/component	A unit produced the asserted result under controlled inputs	Packaging, external providers, end-to-end behavior, or another profile
Synthetic/recorded integration	Protocol/orchestration/error handling against a deterministic double	Genuine inference, provider availability, soak, resilience, release, or readiness
Packaged functional	A named workflow crossed the supported packaged UAR boundary	A model boundary unless genuine model evidence is retained, or any untested path
Real-model functional	A named provider/model performed genuine inference through UAR for the bounded path	Broad quality, every provider, duration resilience, deployment, or another profile
Load/soak/resilience	The named workload exercised its failure model, volume, period, or statistical objective	Inference when traffic uses a double, or claims beyond the workload/environment
Deployment validation	The deployed artifact was reachable and passed named deployment checks	Routine code correctness, inference, or undeployed profiles

Profile and source boundaries

• server-full evidence applies to the named server-full build and configuration.
• minimal evidence is separate; absence of Cedar or packaged UI changes the surface under test.
• embedded-mobile evidence depends on host-provided inference, lifecycle, and persistence and does not inherit server deployment results.

A result also binds to its recorded source SHA. Later source can use the result as history, not as fresh execution evidence. “All profiles” and “the runtime is ready” are aggregate conclusions that require separately stated criteria; one passing row cannot imply them.

Inference boundary

Certifying inference must traverse UAR, use the configured provider path, reach a real loaded model, and return that model's output through UAR. Retain the provider/model identity and, when routing or an agent is under test, the effective policy. Mocked, stubbed, recorded, replayed, hard-coded, or otherwise synthetic responses remain non-certifying diagnostics.

Evidence record minimum

Record the requirement, exact command or interaction, observed result, negative control when fail-closed, source SHA, profile, environment, and explicit limit. If a prerequisite is unavailable, record the claim unverified instead of substituting a weaker test.