“system clock and TLS authenticated package downloads pass controlled tests” addresses operating system clock synchronization was omitted from bootstrap, but the current hardware or infrastructure issue is hardware-specific OS configuration was assumed portable without validation. The option “both platform groups enumerate expected devices and pass baseline I/O tests” is more directly tied to the required outcome.

“DPU initialization and isolation tests pass on the hardened image” addresses OS security settings were applied without hardware-driver compatibility testing, but the current hardware or infrastructure issue is hardware-specific OS configuration was assumed portable without validation. The option “both platform groups enumerate expected devices and pass baseline I/O tests” is more directly tied to the required outcome.

“repeated load tests show latency and power within approved bounds” addresses OS platform power policy conflicts with stable AI compute performance, but the current hardware or infrastructure issue is hardware-specific OS configuration was assumed portable without validation. The option “both platform groups enumerate expected devices and pass baseline I/O tests” is more directly tied to the required outcome.

“validate Pyxis and Enroot device passthrough under Slurm GRES rules” addresses container GPU exposure is not integrated with the scheduler allocation, but the current hardware or infrastructure issue is the GRES device configuration disagrees with installed accelerator enumeration. The option “reconcile Slurm GRES and permitted GPU device mappings” is more directly tied to the required outcome.

“synchronize BCM node categories and Slurm feature or partition configuration” addresses cluster provisioning attributes are inconsistent with scheduler placement rules, but the current hardware or infrastructure issue is the GRES device configuration disagrees with installed accelerator enumeration. The option “reconcile Slurm GRES and permitted GPU device mappings” is more directly tied to the required outcome.

“apply supported cgroup or device isolation in Slurm and its container integration” addresses gPU device access is not constrained to the scheduler allocation, but the current hardware or infrastructure issue is the GRES device configuration disagrees with installed accelerator enumeration. The option “reconcile Slurm GRES and permitted GPU device mappings” is more directly tied to the required outcome.

“submitted container job reads and writes the expected shared paths” addresses container bind mounts and permitted storage paths are misconfigured, but the current hardware or infrastructure issue is the GRES device configuration disagrees with installed accelerator enumeration. The option “sinfo and a submitted GPU job both expose the approved GPU count” is more directly tied to the required outcome.

“the four-GPU job selects an eligible node or queues correctly” addresses slurm partition and GPU resource constraints were not enforced, but the current hardware or infrastructure issue is the GRES device configuration disagrees with installed accelerator enumeration. The option “sinfo and a submitted GPU job both expose the approved GPU count” is more directly tied to the required outcome.

“nvidia-smi inside the scheduled container shows only the assigned GPU devices” addresses container GPU exposure is not integrated with the scheduler allocation, but the current hardware or infrastructure issue is the GRES device configuration disagrees with installed accelerator enumeration. The option “sinfo and a submitted GPU job both expose the approved GPU count” is more directly tied to the required outcome.

“apply supported cgroup or device isolation in Slurm and its container integration” addresses gPU device access is not constrained to the scheduler allocation, but the current hardware or infrastructure issue is scheduling readiness does not require complete node software acceptance. The option “gate Slurm node availability on validated driver and container plugin health” is more directly tied to the required outcome.

Useful for container GPU exposure is not integrated with the scheduler allocation, but it does not remove the specific condition: scheduling readiness does not require complete node software acceptance.

“reconcile Slurm GRES and permitted GPU device mappings” addresses the GRES device configuration disagrees with installed accelerator enumeration, but the current hardware or infrastructure issue is scheduling readiness does not require complete node software acceptance. The option “gate Slurm node availability on validated driver and container plugin health” is more directly tied to the required outcome.

“an sbatch container job starts successfully with the intended image” addresses slurm SPANK plugin integration for Pyxis is incomplete, but the current hardware or infrastructure issue is scheduling readiness does not require complete node software acceptance. The option “test job starts only on fully verified nodes with usable accelerators” is more directly tied to the required outcome.

“a process cannot access a GPU not granted to its job allocation” addresses gPU device access is not constrained to the scheduler allocation, but the current hardware or infrastructure issue is scheduling readiness does not require complete node software acceptance. The option “test job starts only on fully verified nodes with usable accelerators” is more directly tied to the required outcome.

“the four-GPU job selects an eligible node or queues correctly” addresses slurm partition and GPU resource constraints were not enforced, but the current hardware or infrastructure issue is scheduling readiness does not require complete node software acceptance. The option “test job starts only on fully verified nodes with usable accelerators” is more directly tied to the required outcome.

“set appropriate partitions reservations and fair-share limits for competing jobs” addresses scheduler queues and limits do not represent workload isolation policy, but the current hardware or infrastructure issue is slurm partition and GPU resource constraints were not enforced. The option “configure consistent GRES and scheduling constraints for GPU requests” is more directly tied to the required outcome.

“validate Pyxis and Enroot device passthrough under Slurm GRES rules” addresses container GPU exposure is not integrated with the scheduler allocation, but the current hardware or infrastructure issue is slurm partition and GPU resource constraints were not enforced. The option “configure consistent GRES and scheduling constraints for GPU requests” is more directly tied to the required outcome.

“reconcile Slurm GRES and permitted GPU device mappings” addresses the GRES device configuration disagrees with installed accelerator enumeration, but the current hardware or infrastructure issue is slurm partition and GPU resource constraints were not enforced. The option “configure consistent GRES and scheduling constraints for GPU requests” is more directly tied to the required outcome.

“an sbatch container job starts successfully with the intended image” addresses slurm SPANK plugin integration for Pyxis is incomplete, but the current hardware or infrastructure issue is slurm partition and GPU resource constraints were not enforced. The option “the four-GPU job selects an eligible node or queues correctly” is more directly tied to the required outcome.

“submitted container job reads and writes the expected shared paths” addresses container bind mounts and permitted storage paths are misconfigured, but the current hardware or infrastructure issue is slurm partition and GPU resource constraints were not enforced. The option “the four-GPU job selects an eligible node or queues correctly” is more directly tied to the required outcome.

“sinfo and a submitted GPU job both expose the approved GPU count” addresses the GRES device configuration disagrees with installed accelerator enumeration, but the current hardware or infrastructure issue is slurm partition and GPU resource constraints were not enforced. The option “the four-GPU job selects an eligible node or queues correctly” is more directly tied to the required outcome.

“validate Pyxis and Enroot device passthrough under Slurm GRES rules” addresses container GPU exposure is not integrated with the scheduler allocation, but the current hardware or infrastructure issue is slurm SPANK plugin integration for Pyxis is incomplete. The option “install and validate matching Pyxis plugin and Enroot versions across scheduler nodes” is more directly tied to the required outcome.

“gate Slurm node availability on validated driver and container plugin health” addresses scheduling readiness does not require complete node software acceptance, but the current hardware or infrastructure issue is slurm SPANK plugin integration for Pyxis is incomplete. The option “install and validate matching Pyxis plugin and Enroot versions across scheduler nodes” is more directly tied to the required outcome.

“define approved Enroot bind paths for checkpoint and dataset mounts” addresses container bind mounts and permitted storage paths are misconfigured, but the current hardware or infrastructure issue is slurm SPANK plugin integration for Pyxis is incomplete. The option “install and validate matching Pyxis plugin and Enroot versions across scheduler nodes” is more directly tied to the required outcome.

“test job starts only on fully verified nodes with usable accelerators” addresses scheduling readiness does not require complete node software acceptance, but the current hardware or infrastructure issue is slurm SPANK plugin integration for Pyxis is incomplete. The option “an sbatch container job starts successfully with the intended image” is more directly tied to the required outcome.

“submitted container job reads and writes the expected shared paths” addresses container bind mounts and permitted storage paths are misconfigured, but the current hardware or infrastructure issue is slurm SPANK plugin integration for Pyxis is incomplete. The option “an sbatch container job starts successfully with the intended image” is more directly tied to the required outcome.

“the four-GPU job selects an eligible node or queues correctly” addresses slurm partition and GPU resource constraints were not enforced, but the current hardware or infrastructure issue is slurm SPANK plugin integration for Pyxis is incomplete. The option “an sbatch container job starts successfully with the intended image” is more directly tied to the required outcome.

Useful for Slurm partition and GPU resource constraints were not enforced, but it does not remove the specific condition: container bind mounts and permitted storage paths are misconfigured.

“validate Pyxis and Enroot device passthrough under Slurm GRES rules” addresses container GPU exposure is not integrated with the scheduler allocation, but the current hardware or infrastructure issue is container bind mounts and permitted storage paths are misconfigured. The option “define approved Enroot bind paths for checkpoint and dataset mounts” is more directly tied to the required outcome.

“set appropriate partitions reservations and fair-share limits for competing jobs” addresses scheduler queues and limits do not represent workload isolation policy, but the current hardware or infrastructure issue is container bind mounts and permitted storage paths are misconfigured. The option “define approved Enroot bind paths for checkpoint and dataset mounts” is more directly tied to the required outcome.

“sinfo and a submitted GPU job both expose the approved GPU count” addresses the GRES device configuration disagrees with installed accelerator enumeration, but the current hardware or infrastructure issue is container bind mounts and permitted storage paths are misconfigured. The option “submitted container job reads and writes the expected shared paths” is more directly tied to the required outcome.

“an sbatch container job starts successfully with the intended image” addresses slurm SPANK plugin integration for Pyxis is incomplete, but the current hardware or infrastructure issue is container bind mounts and permitted storage paths are misconfigured. The option “submitted container job reads and writes the expected shared paths” is more directly tied to the required outcome.

“nvidia-smi inside the scheduled container shows only the assigned GPU devices” addresses container GPU exposure is not integrated with the scheduler allocation, but the current hardware or infrastructure issue is container bind mounts and permitted storage paths are misconfigured. The option “submitted container job reads and writes the expected shared paths” is more directly tied to the required outcome.