“stress the affected GPU and correlate throttle reasons with system telemetry” addresses gPU cooling or power behavior degrades only under sustained compute duty, but the current hardware or infrastructure issue is aggregate power and thermal behavior across all GPUs was never validated. The option “run simultaneous multi-GPU stress at the approved concurrency” is more directly tied to the required outcome.

Useful for idle health checks failed to exercise sustained compute and power load, but it does not remove the specific condition: aggregate power and thermal behavior across all GPUs was never validated.

“monitor NIC links storage and GPU telemetry during maximum approved node stress” addresses system-level concurrent device behavior is not included in single-node acceptance, but the current hardware or infrastructure issue is aggregate power and thermal behavior across all GPUs was never validated. The option “run simultaneous multi-GPU stress at the approved concurrency” is more directly tied to the required outcome.

“network interfaces remain stable while full node load stays in limits” addresses system-level concurrent device behavior is not included in single-node acceptance, but the current hardware or infrastructure issue is aggregate power and thermal behavior across all GPUs was never validated. The option “aggregate load stays inside rack power and device thermal limits” is more directly tied to the required outcome.

“reruns under identical settings yield comparable measured load and error trends” addresses inconsistent stress conditions invalidate benchmark comparison, but the current hardware or infrastructure issue is aggregate power and thermal behavior across all GPUs was never validated. The option “aggregate load stays inside rack power and device thermal limits” is more directly tied to the required outcome.

“repeated full-duration stress passes without reboot ECC or power alarms” addresses short smoke testing missed a load-dependent power or cooling instability, but the current hardware or infrastructure issue is aggregate power and thermal behavior across all GPUs was never validated. The option “aggregate load stays inside rack power and device thermal limits” is more directly tied to the required outcome.

“run a qualified all-GPU stress profile with temperature and Xid telemetry” addresses idle health checks failed to exercise sustained compute and power load, but the current hardware or infrastructure issue is gPU cooling or power behavior degrades only under sustained compute duty. The option “stress the affected GPU and correlate throttle reasons with system telemetry” is more directly tied to the required outcome.

“monitor NIC links storage and GPU telemetry during maximum approved node stress” addresses system-level concurrent device behavior is not included in single-node acceptance, but the current hardware or infrastructure issue is gPU cooling or power behavior degrades only under sustained compute duty. The option “stress the affected GPU and correlate throttle reasons with system telemetry” is more directly tied to the required outcome.

“extend bounded single-node compute and memory stress under monitored conditions” addresses short smoke testing missed a load-dependent power or cooling instability, but the current hardware or infrastructure issue is gPU cooling or power behavior degrades only under sustained compute duty. The option “stress the affected GPU and correlate throttle reasons with system telemetry” is more directly tied to the required outcome.

“all accelerators sustain the accepted workload duration without critical errors” addresses idle health checks failed to exercise sustained compute and power load, but the current hardware or infrastructure issue is gPU cooling or power behavior degrades only under sustained compute duty. The option “sampled clocks remain above the approved floor without unexpected throttle flags” is more directly tied to the required outcome.

“repeated full-duration stress passes without reboot ECC or power alarms” addresses short smoke testing missed a load-dependent power or cooling instability, but the current hardware or infrastructure issue is gPU cooling or power behavior degrades only under sustained compute duty. The option “sampled clocks remain above the approved floor without unexpected throttle flags” is more directly tied to the required outcome.

“network interfaces remain stable while full node load stays in limits” addresses system-level concurrent device behavior is not included in single-node acceptance, but the current hardware or infrastructure issue is gPU cooling or power behavior degrades only under sustained compute duty. The option “sampled clocks remain above the approved floor without unexpected throttle flags” is more directly tied to the required outcome.

Useful for system-level concurrent device behavior is not included in single-node acceptance, but it does not remove the specific condition: correct numerical outputs do not negate accumulating hardware memory faults.

“run a qualified all-GPU stress profile with temperature and Xid telemetry” addresses idle health checks failed to exercise sustained compute and power load, but the current hardware or infrastructure issue is correct numerical outputs do not negate accumulating hardware memory faults. The option “investigate ECC telemetry and isolate defective memory during a supported stress run” is more directly tied to the required outcome.

“stress the affected GPU and correlate throttle reasons with system telemetry” addresses gPU cooling or power behavior degrades only under sustained compute duty, but the current hardware or infrastructure issue is correct numerical outputs do not negate accumulating hardware memory faults. The option “investigate ECC telemetry and isolate defective memory during a supported stress run” is more directly tied to the required outcome.

“reruns under identical settings yield comparable measured load and error trends” addresses inconsistent stress conditions invalidate benchmark comparison, but the current hardware or infrastructure issue is correct numerical outputs do not negate accumulating hardware memory faults. The option “uncorrectable ECC counters stay clear and error trend meets acceptance rules” is more directly tied to the required outcome.

“stress traces show within-spec operation and relevant workload stability” addresses stress profile does not represent a safe supported commissioning envelope, but the current hardware or infrastructure issue is correct numerical outputs do not negate accumulating hardware memory faults. The option “uncorrectable ECC counters stay clear and error trend meets acceptance rules” is more directly tied to the required outcome.

“repeated full-duration stress passes without reboot ECC or power alarms” addresses short smoke testing missed a load-dependent power or cooling instability, but the current hardware or infrastructure issue is correct numerical outputs do not negate accumulating hardware memory faults. The option “uncorrectable ECC counters stay clear and error trend meets acceptance rules” is more directly tied to the required outcome.

“standardize workload mix duration power cap and monitoring for repeatable tests” addresses inconsistent stress conditions invalidate benchmark comparison, but the current hardware or infrastructure issue is stress profile does not represent a safe supported commissioning envelope. The option “apply vendor-qualified stress parameters within site power and cooling constraints” is more directly tied to the required outcome.

“monitor NIC links storage and GPU telemetry during maximum approved node stress” addresses system-level concurrent device behavior is not included in single-node acceptance, but the current hardware or infrastructure issue is stress profile does not represent a safe supported commissioning envelope. The option “apply vendor-qualified stress parameters within site power and cooling constraints” is more directly tied to the required outcome.

Useful for correct numerical outputs do not negate accumulating hardware memory faults, but it does not remove the specific condition: stress profile does not represent a safe supported commissioning envelope.

“reruns under identical settings yield comparable measured load and error trends” addresses inconsistent stress conditions invalidate benchmark comparison, but the current hardware or infrastructure issue is stress profile does not represent a safe supported commissioning envelope. The option “stress traces show within-spec operation and relevant workload stability” is more directly tied to the required outcome.

“uncorrectable ECC counters stay clear and error trend meets acceptance rules” addresses correct numerical outputs do not negate accumulating hardware memory faults, but the current hardware or infrastructure issue is stress profile does not represent a safe supported commissioning envelope. The option “stress traces show within-spec operation and relevant workload stability” is more directly tied to the required outcome.

“aggregate load stays inside rack power and device thermal limits” addresses aggregate power and thermal behavior across all GPUs was never validated, but the current hardware or infrastructure issue is stress profile does not represent a safe supported commissioning envelope. The option “stress traces show within-spec operation and relevant workload stability” is more directly tied to the required outcome.

“investigate ECC telemetry and isolate defective memory during a supported stress run” addresses correct numerical outputs do not negate accumulating hardware memory faults, but the current hardware or infrastructure issue is burn-in evidence cannot be attributed to the replaced hardware. The option “bind stress artifacts to GPU UUID PCI slot and chassis identity” is more directly tied to the required outcome.

Useful for GPU cooling or power behavior degrades only under sustained compute duty, but it does not remove the specific condition: burn-in evidence cannot be attributed to the replaced hardware.

“run a qualified all-GPU stress profile with temperature and Xid telemetry” addresses idle health checks failed to exercise sustained compute and power load, but the current hardware or infrastructure issue is burn-in evidence cannot be attributed to the replaced hardware. The option “bind stress artifacts to GPU UUID PCI slot and chassis identity” is more directly tied to the required outcome.

“all accelerators sustain the accepted workload duration without critical errors” addresses idle health checks failed to exercise sustained compute and power load, but the current hardware or infrastructure issue is burn-in evidence cannot be attributed to the replaced hardware. The option “results trace directly to the installed replacement board” is more directly tied to the required outcome.

“sampled clocks remain above the approved floor without unexpected throttle flags” addresses gPU cooling or power behavior degrades only under sustained compute duty, but the current hardware or infrastructure issue is burn-in evidence cannot be attributed to the replaced hardware. The option “results trace directly to the installed replacement board” is more directly tied to the required outcome.

“uncorrectable ECC counters stay clear and error trend meets acceptance rules” addresses correct numerical outputs do not negate accumulating hardware memory faults, but the current hardware or infrastructure issue is burn-in evidence cannot be attributed to the replaced hardware. The option “results trace directly to the installed replacement board” is more directly tied to the required outcome.

“investigate ECC telemetry and isolate defective memory during a supported stress run” addresses correct numerical outputs do not negate accumulating hardware memory faults, but the current hardware or infrastructure issue is system-level concurrent device behavior is not included in single-node acceptance. The option “monitor NIC links storage and GPU telemetry during maximum approved node stress” is more directly tied to the required outcome.

“stress the affected GPU and correlate throttle reasons with system telemetry” addresses gPU cooling or power behavior degrades only under sustained compute duty, but the current hardware or infrastructure issue is system-level concurrent device behavior is not included in single-node acceptance. The option “monitor NIC links storage and GPU telemetry during maximum approved node stress” is more directly tied to the required outcome.

Useful for aggregate power and thermal behavior across all GPUs was never validated, but it does not remove the specific condition: system-level concurrent device behavior is not included in single-node acceptance.