Physical workstation acceptance

Every item below is pending. Record date, firmware settings/version, Tacid commit/generation, kernel/driver/CUDA versions, commands, logs and observed failures. A fixture, evaluated Nix system or emulated VM is not PCIe passthrough evidence. Follow the preparation guide first. Keep a recovery keyboard, working boot generation and live medium available. Neither SSD is modified by this checklist unless you explicitly carry out the separate installation plan.

1. Motherboard boot and unlock. With both RTX cards installed and monitor on motherboard output, cold boot and inspect UEFI, boot menu, Plymouth and LUKS passphrase/TPM PIN prompt. Test passphrase fallback and keyboard layout. Save photographs/notes; verify Secure Boot only through the existing enrollment path.
2. Radeon desktop. Log in to the existing Hyprland Tacid desktop. Check readlink -f /dev/dri/tacid-display, lspci -nnk, hyprctl monitors and the actual compositor DRM device usage. Verify amdgpu firmware/logs and display resolution/refresh. Removing NVIDIA from the display role must not hide unlock.
3. Two individual Linux GPUs. In Linux AI mode run nvidia-smi -L and nvidia-smi --query-gpu=name,pci.bus_id,memory.total,driver_version --format=csv. Expect two distinct RTX 5080 devices, about 16 GB each. Record bus IDs/UUIDs.
4. CUDA and distributed inference. From a pinned CUDA 12.9 development shell (owner-approved CUDA packages and compiler), build nvcc -arch=sm_120 nix/tests/hardware/cuda-smoke.cu -o /tmp/tacid-cuda-smoke and run /tmp/tacid-cuda-smoke. Check the result on each device individually with CUDA_VISIBLE_DEVICES set to its freshly observed index/UUID, then both. Run the guide’s licensed/checksum-verified GGUF split example and watch nvidia-smi allocations. Record model hash, quantization, disk size, context, parallel slots, per-card weights/cache/buffers, performance and errors. Test under the real tacid-infer unit: inspect its systemd device policy and engine logs, not just a privileged shell. Test one card and zero cards after VM ownership, with an appropriately sized model/CPU fallback. Confirm class grants after restart.
5. Topology. Record lspci -nnk and every link under /sys/kernel/iommu_groups. Match each RTX GPU and HDMI/DP audio function to physical slots and complete groups. Refuse host-critical shared groups. Check /dev/kvm, SVM flags (lscpu), AMD-Vi boot logs, storage/controller/NIC IDs and exact NVMe serials. Do not substitute IDs from fixtures or enable ACS override.
6. Virtual Windows installation. Install Windows 11 Pro using your ISO/license in windows11 through the virtual console, with OVMF, emulated TPM 2.0, VirtIO disk/network, NAT and guest drivers. Check clean shutdown/restart and virtual console input/audio. Reboot the host and switch generation; confirm domain, disk data, NVRAM and vTPM identity/state survive. Save offline backups together.
7. One-card passthrough. Review/enable actual groups in owner configuration. Start a deliberate CUDA job and prove VM start is refused without killing it; finish it, stop inference, prevent new jobs and assign one complete group. Verify Windows NVIDIA driver and actual RTX physical display output on the second monitor input. Verify Radeon remains usable and Linux CUDA works on the retained card after a service restart. Check status launcher against sysfs/logs.
8. Optional two-card Windows Compute. Shut Windows down before changing XML. Stop all GPU users, assign both complete RTX groups and start Windows. Check both devices individually in Windows and Radeon desktop continuity. Linux CUDA has zero RTX devices in this mode; check CPU fallback rather than GPU inference.
9. Return and repeat. Cleanly shut Windows down; verify NVIDIA binding, nvidia-smi, smoke test and real inference after every return. Repeat at least five start/stop cycles and one cold boot for each accepted mode, checking logs for reset/IOMMU errors. Until demonstrated, return without reboot remains unproven. Record exact recovery actions for failures; never silently force-reset an in-use card. Keep Windows autostart disabled.
10. Peripherals and monitor. Test physical monitor input/KVM switching, dedicated keyboard/mouse and explicit audio route. Keep host recovery input. Hyprland workspace switching must not be described as monitor input switching. Recheck virtual console access when monitor is on Windows output.
11. Recovery. Exercise TTY, opt-in key-only SSH from another machine, previous boot generation and live-medium repair without formatting. Simulate a bad graphics declaration through a staged generation only when a known-good one is preserved; prove Radeon/TTY recovery, restore driver and check both cards.

Do not promote this machine to physically supported until the relevant evidence is recorded. Check off only what was actually observed, including failed transitions.