Native Qualification Run Sequence

Native Qualification Run Sequence

1. Before issuing a permit, build the exact Actor VSIX and NSIS payload, verify source revision, vendor SHA-256 values, publisher signature, internal-NVMe capacity, Guest Additions capability, and all required local ports. A missing precondition stops before VM creation.
2. Compile the host-bound one-use permit, create exactly one disposable native Windows VM on internal NVMe, then independently verify its UUID, resource profile, config/disk paths, powered-on state, NAT rules, and loopback-only host forwards before sealing the run grant.
3. Attach only the run's vendor artifact directory as a transient read-only VirtualBox share. Run the NSIS installer as the interactive user with RequestExecutionLevel user; verify every payload, persist logs and exit status, remove the share, then independently inspect the installed extension version/hash, capture tools, hidden UI/relay tasks, and exact firewall rule.
4. Start the seeded synthetic UI in the guest desktop and issue the Actor desktop-capture command. Discover one opaque capture UUID through Actor resources/list; admit only the session bound to the sealed VM and pinned Actor contract.
5. While the capture reports capturing, route the exact approved benchmark-series call, poll bounded resources, verify the frame hash/PNG and sequence, and stream that frame to loopback Ollama with the approved digest, CPU settings, response schema, and finite deadline.
6. Require the model choice to reference that exact capture/frame and contain the observed reading/limit. The broker accepts only one of the two pre-authorized synthetic targets; the acknowledgement is recorded as task acceptance, not correctness. After that action returns, issue the one capture-stop command and drain final Actor resources.
7. Finalize only after the Actor reports stopped, all admitted frame/sample watermarks are drained, and every frame is verified. Record separate grounding, action, task, and monotonic timing results; redact raw images and model prose; issue the exact one-shot teardown target.
8. Repeat using pre-registered seeds in all three modes. Keep calibration runs separate from qualification runs; freeze numeric thresholds only through REQ-125 after independent review of sample count, uncertainty, configuration, and per-mode distributions.