Q-AXXELE Trace Evidence v3.0.2: Audited External GPU Workload Replay and Pre-Publication Robustness Addendum
Michele Giletto · Zenodo (CERN European Organization for Nuclear Research) · 2026
Q-AXXELE is a solver-neutral decision-evidence architecture for constrained AI-infrastructure scheduling. This publication bundle contains the unchanged audited v3.0.1 external-demand evidence core derived from the Alibaba PAI GPU Cluster Trace 2020 task table, together with an independent publication audit and a v3.0.2 pre-publication robustness addendum. The material includes the frozen replay protocol, executable analysis code, 945 matched route evaluations across 189 day-by-flexibility scenarios, paired-effect outputs, PUE sensitivity, publication figures, errata, machine-readable robustness results, a reproducible robustness script, and SHA-256 manifests. The raw Alibaba trace is not redistributed. The principal public finding is an objective conflict, not a universal performance claim. Under an assumed four-hour start-flexibility counterfactual, an energy-oriented greedy route reduced the modeled accelerator-plus-PUE energy proxy by 0.5976% on average while increasing aggregate requested GPU-equivalent concurrency peak by 17.99% on average. A peak-aware classical route reduced that concurrency peak by 28.49% on average while producing only a 0.0383% average modeled energy-proxy reduction. The simulated-annealing challenger produced no outcome improvement over the energy-oriented greedy route in the released scenarios. No quantum hardware was executed and no quantum advantage is claimed. A post-audit robustness pass found that the four-hour peak conflict is not driven by one trace day. Circular 7-day block-bootstrap intervals remain positive for the PUE-aware peak increase and negative for the peak-aware route; leave-one-day-out means do not change sign. The same directional conflict is visible in a second released congestion proxy, maximum GPU-type capacity utilisation. The nominal device-power coefficients can change modeled-energy magnitude but do not determine the PUE-aware task start ranking or concurrency schedule. Energy, device power, PUE, capacity headroom, and permitted start flexibility are counterfactual model inputs. No wall-plug power, customer telemetry, customer saving, continuous-cluster reconstruction, multi-resolution peak validation, production deployment, or quantum advantage is measured in this release. Continuous carry-in/carry-out replay and 5/15/30-minute peak-resolution testing remain v4 evidence gates.