uat

User Acceptance Testing

Start with acceptance criteria, a production-equivalent public entry point or target host package, environment constraints, credentials, and cost limits. Translate each criterion into an outside-in observable scenario without reading implementation code. Confirm safe drive mechanics, run the narrowest real path fresh, provide one input at a time, observe before continuing, and record interaction, environment, expected result, actual result, and durable evidence. Exercise an adjacent or error path only when it can change acceptance.

Every artifact the drive produces, including screenshots, recordings, logs, probe pages, and captured output, goes in the session scratch directory, never in the working tree of the repository under test. Many capture tools write to the current directory by default, so give each one an explicit absolute output path inside the scratch directory.

Stop when every criterion is passed, failed, or blocked, one critical failure makes further work wasteful, or the stated boundary is reached. Return per-criterion status, exact reproduction, observations, evidence locations, untested scope, and blockers.

Missing runtime, authentication, dependency, safe environment, safe drive mechanics, or host produces blocked, not a pass. Scripted tests support but never replace direct observation of an exercisable surface.

The Real Operation

A surface is exercised only when its production code path runs end to end: every outbound call, every database mutation, every billable side effect a real user triggers. Isolate state where you can (a per-run sandbox directory, an isolated $HOME, a scoped database schema, a test inbox) so the drive does not pollute the developer's real local data, but isolation of state is not the same as faking the operation. Only the local state container is scoped; the services the product is configured for are exercised against their real endpoints.

Mocks, fakes, stubs, fixtures, in-memory adapters, sandbox-provider URLs, dry-run flags, frozen clocks, recorded-cassette replay, and provider-mocking middleware are silent downgrades unless they are the production code path. Inspect the configuration the product runs with (environment variables, config file, resolved client URLs, the credential the runtime picked up) and confirm the drive routes through the same backend a real user's instance would. If it does not, you are exercising a different product.