Testing methodology history
UAR did not stop using small tests when it adopted genuine-model functional verification. The evidence changed because the claim changed. A unit test is appropriate for a pure decision. A recorded provider is useful for deterministic protocol behavior. Neither proves that a configured provider reached a real model and returned inference through the packaged runtime.
Coverage and isolated correctness
ADR-0003 established a 60% local line-coverage target for Rust and React. The retained baseline table is still unpopulated, so the repository does not claim an observed 60% baseline. Its old statement that CI would enforce coverage is superseded: routine verification is local and GitHub Actions are deployment-only.
Coverage answers which lines ran under a suite. Focused unit and component tests answer whether a selected decision produced its asserted result. Neither answers whether the packaged application, external provider, or user workflow worked.
Recorded-provider diagnostics
The 1.0 readiness phase used a deterministic recorded provider for a serial server-full capability matrix. That was valid for the stated authentication, tenancy, skill, and orchestration assertions. Negative controls showed that the fail-closed cases could actually fail. The results transferred to no other profile and did not certify real model inference.
The synthetic-soak failure
A later operational plan spent hours sending deterministic traffic to a local provider double. The run exercised runtime plumbing, then a shutdown test exposed a real deadline defect. But elapsed time did not repair the workload's missing model boundary. The evidence could not support the requested inference or broad readiness claim.
The project therefore rejected duration-only synthetic evidence. A long run now needs a named failure model, traffic-volume target, operating-period target, or statistical detection objective. If it uses a double, it still cannot certify inference.
Bounded real-model closeout
On 2026-08-22, UAR 1.0 closed on five functions—proxy inference, skill activation, knowledge grounding, Kimi k3 configuration/inference, and basic-agent creation/inference—each exercised without and through the shipped UI. The retained evidence identifies genuine models and the effective policy/routing where relevant.
The soak, supply-chain, release-candidate, and GA-publication changes were
cancelled, not called passed. The functional result was local server-full
evidence only. It made no minimal, embedded-mobile, supply-chain, publication,
or universal production-readiness claim.
Current rule
Choose the narrowest test that crosses the boundary named in the requirement. Keep cheap diagnostics for fast fault localization. Add genuine external integration only when the claim reaches that boundary. Retain what the result does not prove as carefully as what it does.