{
  "schema": "aureon.public_benchmark_summary.v1",
  "cohort": "aureon-clock-qualified-20261027-r1",
  "date": "2026-10-27",
  "version": "0.1.0",
  "scope": {
    "model": "TerraMind 1.0 base",
    "task": "prepared real 224x224 S2L2A to synthetic S1GRD",
    "hardware": "RTX 3070 8 GiB, WSL",
    "dependencies": "PyTorch 2.10.0+cu128, TerraTorch 1.2.13, vLLM 0.19.0",
    "precision": "highest FP32; TF32 and AMP off",
    "generation": "8 token steps / 10 diffusion steps",
    "baseline": "TerraTorch + native vLLM with necessary independent TerraMind adapter and corrected generic I/O; explicit batch-one/in-process/0.85 reservation overrides",
    "candidate": "Installed v0.1.0 wheel 0fa5ef2b9440fcc310a17c76f78df52278cbe2d354015be2046a8e4196715c97; resident, reuse=auto, denoiser=graph, full guards, descriptor I/O",
    "measurement": "Full input read/upload through verified saved output, local IPC, serial batch1/FIFO3; not HTTP or provider benchmark",
    "order": "U-A-A-U, reverse burst seed order in second pair",
    "counts": "20 logical jobs + 8 conditioning; 2 sequential and 6 burst observations per arm",
    "parity": "28/28 exact raw/token/post-RNG/COG/grid/mask/provenance; cold full logits; 4 workers close CUDA0/0",
    "clock": "RAW elapsed timing qualified against 1900 Windows QPC samples",
    "cost_rate": "1 THB per machine-hour; normalized, not hosted pricing"
  },
  "measured_wheel_sha256": "0fa5ef2b9440fcc310a17c76f78df52278cbe2d354015be2046a8e4196715c97",
  "selected_local_wheel_sha256": "e09e710c9e9ace17b7bf2214d8c9eff2c4ed605a7d45e8c4e649b672f6f88a64",
  "measures_selected_local_wheel": false,
  "artifact_scope": "Historical original Best comparison. Do not transfer percentages to another wheel, HTTP, scenes or hosted service.",
  "baseline_label": "TerraTorch + native vLLM with an independent TerraMind adapter and disclosed overrides; not an unmodified default end-to-end product",
  "metrics": {
    "upstream": {
      "pooled_burst_jobs_per_second": 1.8667027797327143,
      "sequential_ms": 617.4288775,
      "burst_completion_ms": 1102.199868,
      "warm_THB_per_1000_jobs_at_1_per_hour": 0.14880664495370372
    },
    "aureon": {
      "pooled_burst_jobs_per_second": 2.349801572494419,
      "sequential_ms": 446.3989345,
      "burst_completion_ms": 852.636776,
      "warm_THB_per_1000_jobs_at_1_per_hour": 0.11821329129629628
    }
  },
  "improvements_percent": {
    "sequential_latency_reduction": 27.7003472355405,
    "burst_latency_reduction": 22.6422719912719,
    "throughput_increase": 25.879791791539386,
    "warm_cost_reduction": 20.55913139290625,
    "sampled_RSS_reduction": 12.711891423961053
  },
  "method_notes": {
    "GPU_note": "GPU clocks observed, not locked; first U residency P0, other residencies P2. Both orders improve, including P2-vs-P2. Retain as concise methodology note; it does not negate this observed configuration win and is not a trigger for another benchmark or a release gate.",
    "memory_note": "Peak CUDA allocated nearly unchanged; AUREON reserves 46 MiB more. Sampled RSS 12.71% lower. Do not claim all memory metrics improve.",
    "historical_note": "Do not reuse +41.39% or +38.35% as latest headline; no averaging, pooling, multiplying or retrospective clock rescaling of historical runs.",
    "quality_note": "Exact FP32/reference equality is execution fidelity, not proof that generated SAR is a real observation or more physically accurate."
  },
  "source_handoff_sha256": "b4ec0ad20edaa5200b304cb5d8a21acee9ad67da2c9463b38682e90f025db7c2",
  "source_metrics_sha256": "5f8f8ea5bffda4a7856661c2a6f1f0a9df34863b2d7453a5cc4a3ed06518d527"
}