“measured cold-start performance meets the workload-specific threshold” addresses caching behavior was not controlled in storage acceptance benchmarking, but the current hardware or infrastructure issue is physical storage-path failover has not been validated. The option “controlled single-path failure preserves allowed client data access” is more directly tied to the required outcome.

“observed sample delivery and file-open rates meet the job's target” addresses storage benchmark does not represent workload metadata and small-file behavior, but the current hardware or infrastructure issue is physical storage-path failover has not been validated. The option “controlled single-path failure preserves allowed client data access” is more directly tied to the required outcome.

“aggregate throughput and tail latency meet acceptance at target concurrency” addresses storage concurrency and metadata scaling were not included in acceptance tests, but the current hardware or infrastructure issue is physical storage-path failover has not been validated. The option “controlled single-path failure preserves allowed client data access” is more directly tied to the required outcome.

Useful for storage benchmark does not represent workload metadata and small-file behavior, but it does not remove the specific condition: caching behavior was not controlled in storage acceptance benchmarking.

“benchmark storage from the planned worker count and network paths” addresses storage concurrency and metadata scaling were not included in acceptance tests, but the current hardware or infrastructure issue is caching behavior was not controlled in storage acceptance benchmarking. The option “measure cold and warm data reads separately with documented cache conditions” is more directly tied to the required outcome.

“perform checkpoint write verify and independent restore tests” addresses storage write completion was confused with application-level restore integrity, but the current hardware or infrastructure issue is caching behavior was not controlled in storage acceptance benchmarking. The option “measure cold and warm data reads separately with documented cache conditions” is more directly tied to the required outcome.

“observed sample delivery and file-open rates meet the job's target” addresses storage benchmark does not represent workload metadata and small-file behavior, but the current hardware or infrastructure issue is caching behavior was not controlled in storage acceptance benchmarking. The option “measured cold-start performance meets the workload-specific threshold” is more directly tied to the required outcome.

“aggregate throughput and tail latency meet acceptance at target concurrency” addresses storage concurrency and metadata scaling were not included in acceptance tests, but the current hardware or infrastructure issue is caching behavior was not controlled in storage acceptance benchmarking. The option “measured cold-start performance meets the workload-specific threshold” is more directly tied to the required outcome.

“checkpoint latency remains below the agreed service objective” addresses tail-latency variation is hidden by average throughput measurements, but the current hardware or infrastructure issue is caching behavior was not controlled in storage acceptance benchmarking. The option “measured cold-start performance meets the workload-specific threshold” is more directly tied to the required outcome.

“inject controlled storage faults and validate retry and checkpoint integrity” addresses failure recovery and transactional integrity were omitted from storage testing, but the current hardware or infrastructure issue is test identity differs from real service principal and masks authorization failure. The option “use actual scheduler and container identity for storage access tests” is more directly tied to the required outcome.

Useful for storage concurrency and metadata scaling were not included in acceptance tests, but it does not remove the specific condition: test identity differs from real service principal and masks authorization failure.

“test random metadata-heavy reads under the actual training file distribution” addresses storage benchmark does not represent workload metadata and small-file behavior, but the current hardware or infrastructure issue is test identity differs from real service principal and masks authorization failure. The option “use actual scheduler and container identity for storage access tests” is more directly tied to the required outcome.

“recovered workflow restores last complete checkpoint without corruption” addresses failure recovery and transactional integrity were omitted from storage testing, but the current hardware or infrastructure issue is test identity differs from real service principal and masks authorization failure. The option “intended nonprivileged workload reads and writes approved paths successfully” is more directly tied to the required outcome.

“aggregate throughput and tail latency meet acceptance at target concurrency” addresses storage concurrency and metadata scaling were not included in acceptance tests, but the current hardware or infrastructure issue is test identity differs from real service principal and masks authorization failure. The option “intended nonprivileged workload reads and writes approved paths successfully” is more directly tied to the required outcome.

“observed sample delivery and file-open rates meet the job's target” addresses storage benchmark does not represent workload metadata and small-file behavior, but the current hardware or infrastructure issue is test identity differs from real service principal and masks authorization failure. The option “intended nonprivileged workload reads and writes approved paths successfully” is more directly tied to the required outcome.

“use actual scheduler and container identity for storage access tests” addresses test identity differs from real service principal and masks authorization failure, but the current hardware or infrastructure issue is failure recovery and transactional integrity were omitted from storage testing. The option “inject controlled storage faults and validate retry and checkpoint integrity” is more directly tied to the required outcome.

“benchmark storage from the planned worker count and network paths” addresses storage concurrency and metadata scaling were not included in acceptance tests, but the current hardware or infrastructure issue is failure recovery and transactional integrity were omitted from storage testing. The option “inject controlled storage faults and validate retry and checkpoint integrity” is more directly tied to the required outcome.

“perform checkpoint write verify and independent restore tests” addresses storage write completion was confused with application-level restore integrity, but the current hardware or infrastructure issue is failure recovery and transactional integrity were omitted from storage testing. The option “inject controlled storage faults and validate retry and checkpoint integrity” is more directly tied to the required outcome.

“aggregate throughput and tail latency meet acceptance at target concurrency” addresses storage concurrency and metadata scaling were not included in acceptance tests, but the current hardware or infrastructure issue is failure recovery and transactional integrity were omitted from storage testing. The option “recovered workflow restores last complete checkpoint without corruption” is more directly tied to the required outcome.

“intended nonprivileged workload reads and writes approved paths successfully” addresses test identity differs from real service principal and masks authorization failure, but the current hardware or infrastructure issue is failure recovery and transactional integrity were omitted from storage testing. The option “recovered workflow restores last complete checkpoint without corruption” is more directly tied to the required outcome.

“checkpoint latency remains below the agreed service objective” addresses tail-latency variation is hidden by average throughput measurements, but the current hardware or infrastructure issue is failure recovery and transactional integrity were omitted from storage testing. The option “recovered workflow restores last complete checkpoint without corruption” is more directly tied to the required outcome.

“inspect facility supply temperatures and rack airflow before replacing GPU” addresses inadequate intake cooling is external to the GPU fan modules, but the current hardware or infrastructure issue is gPU diagnostic events were not correlated with the reported workload failures. The option “inspect nvidia-smi DCGM and GPU Xid logs before changing hardware” is more directly tied to the required outcome.

“correlate GPU launch power draw with PSU and feed alert logs” addresses aggregate system power condition is the likely fault domain, but the current hardware or infrastructure issue is gPU diagnostic events were not correlated with the reported workload failures. The option “inspect nvidia-smi DCGM and GPU Xid logs before changing hardware” is more directly tied to the required outcome.

Useful for PCIe slot or electrical signaling fault is blocking reliable device transfers, but it does not remove the specific condition: GPU diagnostic events were not correlated with the reported workload failures.

“repeated DMA tests finish without new uncorrectable PCIe errors” addresses pCIe slot or electrical signaling fault is blocking reliable device transfers, but the current hardware or infrastructure issue is gPU diagnostic events were not correlated with the reported workload failures. The option “fault timestamps identify the affected GPU and post-fix workload runs without new Xids” is more directly tied to the required outcome.

“acceptable inlet and device temperatures return after airflow correction” addresses inadequate intake cooling is external to the GPU fan modules, but the current hardware or infrastructure issue is gPU diagnostic events were not correlated with the reported workload failures. The option “fault timestamps identify the affected GPU and post-fix workload runs without new Xids” is more directly tied to the required outcome.

“an all-GPU stress run has no current limit or PSU fault after correction” addresses aggregate system power condition is the likely fault domain, but the current hardware or infrastructure issue is gPU diagnostic events were not correlated with the reported workload failures. The option “fault timestamps identify the affected GPU and post-fix workload runs without new Xids” is more directly tied to the required outcome.

Useful for inadequate intake cooling is external to the GPU fan modules, but it does not remove the specific condition: chassis cooling fault is causing workload-driven thermal protection.

“correlate AER errors with the device slot and verify hardware seating” addresses pCIe slot or electrical signaling fault is blocking reliable device transfers, but the current hardware or infrastructure issue is chassis cooling fault is causing workload-driven thermal protection. The option “verify fan RPM BMC events and inlet temperatures then restore approved cooling” is more directly tied to the required outcome.

“correlate GPU launch power draw with PSU and feed alert logs” addresses aggregate system power condition is the likely fault domain, but the current hardware or infrastructure issue is chassis cooling fault is causing workload-driven thermal protection. The option “verify fan RPM BMC events and inlet temperatures then restore approved cooling” is more directly tied to the required outcome.

“system remains stable under sustained approved load with both PSUs healthy” addresses power-delivery faults coincide with load rather than application software bugs, but the current hardware or infrastructure issue is chassis cooling fault is causing workload-driven thermal protection. The option “all fans report healthy speed and nodes survive sustained GPU load” is more directly tied to the required outcome.

“acceptable inlet and device temperatures return after airflow correction” addresses inadequate intake cooling is external to the GPU fan modules, but the current hardware or infrastructure issue is chassis cooling fault is causing workload-driven thermal protection. The option “all fans report healthy speed and nodes survive sustained GPU load” is more directly tied to the required outcome.

“NIC enumerates and ports complete full-rate diagnostics after targeted repair” addresses nIC firmware power or PCIe enumeration path may be faulty, but the current hardware or infrastructure issue is chassis cooling fault is causing workload-driven thermal protection. The option “all fans report healthy speed and nodes survive sustained GPU load” is more directly tied to the required outcome.