“cross-rack NCCL bus bandwidth meets the approved topology baseline” addresses end-to-end collective communication path has an unverified performance bottleneck, but the current hardware or infrastructure issue is specific nodes or network paths degrade the larger collective. The option “repaired node set achieves comparable bandwidth across all test subsets” is more directly tied to the required outcome.

“tuned network path meets accepted latency and NCCL bandwidth without persistent congestion” addresses microburst congestion and flow control effects are hidden by averages, but the current hardware or infrastructure issue is specific nodes or network paths degrade the larger collective. The option “repaired node set achieves comparable bandwidth across all test subsets” is more directly tied to the required outcome.

“port counters show both authorized rails carry expected traffic” addresses intended fabric rail balancing is not realized in collective traffic, but the current hardware or infrastructure issue is specific nodes or network paths degrade the larger collective. The option “repaired node set achieves comparable bandwidth across all test subsets” is more directly tied to the required outcome.

“choose and verify intended network interfaces for every NCCL process” addresses communication ranks are bound to nonfunctional or slow fabric adapters, but the current hardware or infrastructure issue is process placement and communication topology are misaligned. The option “map ranks to the intended GPU pod and fabric topology before benchmarking” is more directly tied to the required outcome.

“standardize rank count message sizes and throughput interpretation across runs” addresses collective-size and participant-count effects are ignored in baseline comparisons, but the current hardware or infrastructure issue is process placement and communication topology are misaligned. The option “map ranks to the intended GPU pod and fabric topology before benchmarking” is more directly tied to the required outcome.

“run NCCL tests across the involved racks with per-link congestion monitoring” addresses end-to-end collective communication path has an unverified performance bottleneck, but the current hardware or infrastructure issue is process placement and communication topology are misaligned. The option “map ranks to the intended GPU pod and fabric topology before benchmarking” is more directly tied to the required outcome.

“NCCL logs show correct HCA selection and full-speed all-reduce” addresses communication ranks are bound to nonfunctional or slow fabric adapters, but the current hardware or infrastructure issue is process placement and communication topology are misaligned. The option “alternative validated rank placements show balanced cross-pod throughput” is more directly tied to the required outcome.

“results show matched parameters and explain expected scaling behavior” addresses collective-size and participant-count effects are ignored in baseline comparisons, but the current hardware or infrastructure issue is process placement and communication topology are misaligned. The option “alternative validated rank placements show balanced cross-pod throughput” is more directly tied to the required outcome.

“cross-rack NCCL bus bandwidth meets the approved topology baseline” addresses end-to-end collective communication path has an unverified performance bottleneck, but the current hardware or infrastructure issue is process placement and communication topology are misaligned. The option “alternative validated rank placements show balanced cross-pod throughput” is more directly tied to the required outcome.

“correlate NCCL failure timing with per-GPU Xid and interconnect events” addresses node GPU faults were not distinguished from network communication failure, but the current hardware or infrastructure issue is short network benchmarks failed to detect a time-dependent collective fault. The option “run prolonged NCCL collective loops with telemetry and timeout thresholds” is more directly tied to the required outcome.

Useful for fabric congestion control was never validated at production concurrency, but it does not remove the specific condition: short network benchmarks failed to detect a time-dependent collective fault.

“monitor cooling throttles and per-GPU clocks together with NCCL progress” addresses fabric burn-in also exercises node thermal and compute stability, but the current hardware or infrastructure issue is short network benchmarks failed to detect a time-dependent collective fault. The option “run prolonged NCCL collective loops with telemetry and timeout thresholds” is more directly tied to the required outcome.

“isolated node repair allows collective burn-in without GPU errors” addresses node GPU faults were not distinguished from network communication failure, but the current hardware or infrastructure issue is short network benchmarks failed to detect a time-dependent collective fault. The option “full burn-in interval completes without hang or unrecoverable link errors” is more directly tied to the required outcome.

“sustained measurements remain above the prescribed lower bound throughout” addresses peak sample does not demonstrate stable collective performance over time, but the current hardware or infrastructure issue is short network benchmarks failed to detect a time-dependent collective fault. The option “full burn-in interval completes without hang or unrecoverable link errors” is more directly tied to the required outcome.

“supported failover preserves expected job behavior with clear rail alarms” addresses redundancy failover and fault visibility are untested under sustained collectives, but the current hardware or infrastructure issue is short network benchmarks failed to detect a time-dependent collective fault. The option “full burn-in interval completes without hang or unrecoverable link errors” is more directly tied to the required outcome.

“inject controlled single-rail interruption during monitored NCCL burn-in” addresses redundancy failover and fault visibility are untested under sustained collectives, but the current hardware or infrastructure issue is passing collectives can conceal degrading fabric signal margins. The option “collect network FEC retransmit and error trends during NCCL burn-in” is more directly tied to the required outcome.

“use time-series NCCL throughput and reliability checks across a fixed burn-in period” addresses peak sample does not demonstrate stable collective performance over time, but the current hardware or infrastructure issue is passing collectives can conceal degrading fabric signal margins. The option “collect network FEC retransmit and error trends during NCCL burn-in” is more directly tied to the required outcome.

“pin NCCL settings GPU count and message profile for accepted burn-in runs” addresses changing benchmark parameters invalidates longitudinal stability comparison, but the current hardware or infrastructure issue is passing collectives can conceal degrading fabric signal margins. The option “collect network FEC retransmit and error trends during NCCL burn-in” is more directly tied to the required outcome.

“supported failover preserves expected job behavior with clear rail alarms” addresses redundancy failover and fault visibility are untested under sustained collectives, but the current hardware or infrastructure issue is passing collectives can conceal degrading fabric signal margins. The option “long-run error growth remains within published site limits” is more directly tied to the required outcome.

“full burn-in interval completes without hang or unrecoverable link errors” addresses short network benchmarks failed to detect a time-dependent collective fault, but the current hardware or infrastructure issue is passing collectives can conceal degrading fabric signal margins. The option “long-run error growth remains within published site limits” is more directly tied to the required outcome.

“all racks sustain accepted bandwidth and service stability during duration” addresses fabric congestion control was never validated at production concurrency, but the current hardware or infrastructure issue is passing collectives can conceal degrading fabric signal margins. The option “long-run error growth remains within published site limits” is more directly tied to the required outcome.

“run prolonged NCCL collective loops with telemetry and timeout thresholds” addresses short network benchmarks failed to detect a time-dependent collective fault, but the current hardware or infrastructure issue is fabric congestion control was never validated at production concurrency. The option “run rack-spanning NCCL burn-in at planned scale with switch queue metrics” is more directly tied to the required outcome.

Useful for redundancy failover and fault visibility are untested under sustained collectives, but it does not remove the specific condition: fabric congestion control was never validated at production concurrency.

“collect network FEC retransmit and error trends during NCCL burn-in” addresses passing collectives can conceal degrading fabric signal margins, but the current hardware or infrastructure issue is fabric congestion control was never validated at production concurrency. The option “run rack-spanning NCCL burn-in at planned scale with switch queue metrics” is more directly tied to the required outcome.

“repaired cooling maintains normal clocks and stable collective iteration time” addresses fabric burn-in also exercises node thermal and compute stability, but the current hardware or infrastructure issue is fabric congestion control was never validated at production concurrency. The option “all racks sustain accepted bandwidth and service stability during duration” is more directly tied to the required outcome.

“full burn-in interval completes without hang or unrecoverable link errors” addresses short network benchmarks failed to detect a time-dependent collective fault, but the current hardware or infrastructure issue is fabric congestion control was never validated at production concurrency. The option “all racks sustain accepted bandwidth and service stability during duration” is more directly tied to the required outcome.

“long-run error growth remains within published site limits” addresses passing collectives can conceal degrading fabric signal margins, but the current hardware or infrastructure issue is fabric congestion control was never validated at production concurrency. The option “all racks sustain accepted bandwidth and service stability during duration” is more directly tied to the required outcome.

“collect network FEC retransmit and error trends during NCCL burn-in” addresses passing collectives can conceal degrading fabric signal margins, but the current hardware or infrastructure issue is peak sample does not demonstrate stable collective performance over time. The option “use time-series NCCL throughput and reliability checks across a fixed burn-in period” is more directly tied to the required outcome.

“pin NCCL settings GPU count and message profile for accepted burn-in runs” addresses changing benchmark parameters invalidates longitudinal stability comparison, but the current hardware or infrastructure issue is peak sample does not demonstrate stable collective performance over time. The option “use time-series NCCL throughput and reliability checks across a fixed burn-in period” is more directly tied to the required outcome.

“run prolonged NCCL collective loops with telemetry and timeout thresholds” addresses short network benchmarks failed to detect a time-dependent collective fault, but the current hardware or infrastructure issue is peak sample does not demonstrate stable collective performance over time. The option “use time-series NCCL throughput and reliability checks across a fixed burn-in period” is more directly tied to the required outcome.

“full burn-in interval completes without hang or unrecoverable link errors” addresses short network benchmarks failed to detect a time-dependent collective fault, but the current hardware or infrastructure issue is peak sample does not demonstrate stable collective performance over time. The option “sustained measurements remain above the prescribed lower bound throughout” is more directly tied to the required outcome.

“isolated node repair allows collective burn-in without GPU errors” addresses node GPU faults were not distinguished from network communication failure, but the current hardware or infrastructure issue is peak sample does not demonstrate stable collective performance over time. The option “sustained measurements remain above the prescribed lower bound throughout” is more directly tied to the required outcome.

“long-run error growth remains within published site limits” addresses passing collectives can conceal degrading fabric signal margins, but the current hardware or infrastructure issue is peak sample does not demonstrate stable collective performance over time. The option “sustained measurements remain above the prescribed lower bound throughout” is more directly tied to the required outcome.