{
  "workspace": ":4096:8",
  "action": "measure",
  "scope": "task",
  "command": [
    "/usr/local/bin/python",
    "/workspace/cache/fixture.py",
    "--input",
    "/tmp/cache-run/input.npz",
    "--config",
    "/tmp/cache-run/config.json",
    "--output",
    "/tmp/cache-run/result.json",
    "--mode",
    "measure",
    "--scope",
    "task",
    "--repeats",
    "16",
    "--energy"
  ],
  "returncode": 0,
  "stdout": "Energy trial 0 1059.7543262993806 mJ/task\nEnergy trial 1 1000.6192837807083 mJ/task\nEnergy trial 2 948.5655289801525 mJ/task\nCompleted measure/task: /tmp/cache-run/result.json\n",
  "stderr": "",
  "capabilities": {
    "ncu_version": "NVIDIA (R) Nsight Compute Command Line Profiler\nCopyright (c) 2018-2025 NVIDIA Corporation\nVersion 2025.1.1.0 (build 35528883) (public-release)\n"
  },
  "finished_at_utc": "2026-09-14T23:53:15.468625+00:00",
  "input_sha256": {
    "train_images": "aa68fb7d65b7dcd9c91fd62e48b42ed6a60f7e1fad7c53b1b0ec2c73a5a7c5d1",
    "train_labels": "72abc25a7a08e37d3ae171b30b3e5617ce288cafab5f4df112b352b93f2133fb",
    "test_images": "78aa1bca675869c1e08ba944b99d45dce473bee70d893240be8d733acc6d0144"
  },
  "result": {
    "schema_version": 1,
    "completed_at_utc": "2026-09-14T23:53:14.951098+00:00",
    "mode": "measure",
    "requested_scope": "task",
    "config": {
      "alpha": 0.5,
      "batch_size": 128,
      "momentum": 0.9,
      "weight_decay": 0.0,
      "seed": 11,
      "depth": 2,
      "epochs": 2,
      "learning_rate": 0.1
    },
    "input_archive_sha256": "5ed99d543985278c222ae2e9adea0460d530c1985faae1ddc874e126bedf1131",
    "config_file_sha256": "a4cf957d95e7b2d33db7a3121a0ae00609cd278f79026e4436549103647c618f",
    "source_sha256": {
      "fixture.py": "18370a6bd4ec0e92d446af608b068a86d20d898b293a0edfae753ebb103f2b6a",
      "learner.py": "c395a8b7bc2951e3fdab5c5f5edf5ad2d5357de4e9170e52b988f4312e7169c5",
      "energy.py": "39816c3ba6bfc726f77f45094a8eba61e3fa48e50b93300129ef16878b9beb7a"
    },
    "cublas_workspace_config": ":4096:8",
    "environment_was_set_externally": true,
    "software": {
      "python": "3.11.15",
      "torch": "2.5.1+cu124",
      "cuda": "12.4",
      "numpy": "2.2.6"
    },
    "gpu": "NVIDIA A100-SXM4-40GB",
    "cuda_device_index": 0,
    "memory": {
      "allocated_before_prepare_bytes": 0,
      "reserved_before_prepare_bytes": 0,
      "allocated_after_prepare_bytes": 81084416,
      "prepare_and_invoke_peak_allocated_bytes": 82509824,
      "allocated_after_invoke_bytes": 81084416,
      "reserved_after_invoke_bytes": 96468992,
      "prepare_and_invoke_peak_reserved_bytes": 96468992,
      "scope": "Fresh task preparation including CUDA graph capture plus one complete reset/train/predict invocation; tensor allocator only, excluding driver/context. Recorded before output diagnostics and profiling warmup.",
      "added_peak_over_entry_allocated_bytes": 82509824
    },
    "buffer_inventory": {
      "named_tensors": [
        {
          "name": "parameters.weights.0",
          "storage_id": 0,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "gradients.weights.0",
          "storage_id": 1,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "parameters.weights.1",
          "storage_id": 2,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "gradients.weights.1",
          "storage_id": 3,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "parameters.weights.2",
          "storage_id": 4,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "gradients.weights.2",
          "storage_id": 5,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "parameters.weights.3",
          "storage_id": 6,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "gradients.weights.3",
          "storage_id": 7,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "parameters.head.weight",
          "storage_id": 8,
          "shape": [
            10,
            82
          ],
          "stride": [
            82,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 3280
        },
        {
          "name": "gradients.head.weight",
          "storage_id": 9,
          "shape": [
            10,
            82
          ],
          "stride": [
            82,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 3280
        },
        {
          "name": "parameters.head.bias",
          "storage_id": 10,
          "shape": [
            10
          ],
          "stride": [
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 40
        },
        {
          "name": "gradients.head.bias",
          "storage_id": 11,
          "shape": [
            10
          ],
          "stride": [
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 40
        },
        {
          "name": "initial.0",
          "storage_id": 12,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "initial.1",
          "storage_id": 13,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "initial.2",
          "storage_id": 14,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "initial.3",
          "storage_id": 15,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "initial.4",
          "storage_id": 16,
          "shape": [
            10,
            82
          ],
          "stride": [
            82,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 3280
        },
        {
          "name": "initial.5",
          "storage_id": 17,
          "shape": [
            10
          ],
          "stride": [
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 40
        },
        {
          "name": "velocity.0",
          "storage_id": 18,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "velocity.1",
          "storage_id": 19,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "velocity.2",
          "storage_id": 20,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "velocity.3",
          "storage_id": 21,
          "shape": [
            41,
            41
          ],
          "stride": [
            41,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 6724
        },
        {
          "name": "velocity.4",
          "storage_id": 22,
          "shape": [
            10,
            82
          ],
          "stride": [
            82,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 3280
        },
        {
          "name": "velocity.5",
          "storage_id": 23,
          "shape": [
            10
          ],
          "stride": [
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 40
        },
        {
          "name": "raw",
          "storage_id": 24,
          "shape": [
            10000,
            81
          ],
          "stride": [
            81,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 3240000
        },
        {
          "name": "raw_query",
          "storage_id": 25,
          "shape": [
            10000,
            81
          ],
          "stride": [
            81,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 3240000
        },
        {
          "name": "labels",
          "storage_id": 26,
          "shape": [
            10000
          ],
          "stride": [
            1
          ],
          "dtype": "torch.int64",
          "storage_offset": 0,
          "logical_bytes": 80000
        },
        {
          "name": "x",
          "storage_id": 27,
          "shape": [
            10000,
            82
          ],
          "stride": [
            82,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 3280000
        },
        {
          "name": "q",
          "storage_id": 28,
          "shape": [
            10000,
            82
          ],
          "stride": [
            82,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 3280000
        },
        {
          "name": "orders",
          "storage_id": 29,
          "shape": [
            2,
            10000
          ],
          "stride": [
            10000,
            1
          ],
          "dtype": "torch.int64",
          "storage_offset": 0,
          "logical_bytes": 160000
        },
        {
          "name": "order",
          "storage_id": 30,
          "shape": [
            10000
          ],
          "stride": [
            1
          ],
          "dtype": "torch.int64",
          "storage_offset": 0,
          "logical_bytes": 80000
        },
        {
          "name": "predictions",
          "storage_id": 31,
          "shape": [
            10000
          ],
          "stride": [
            1
          ],
          "dtype": "torch.int64",
          "storage_offset": 0,
          "logical_bytes": 80000
        },
        {
          "name": "scores",
          "storage_id": 32,
          "shape": [
            10000,
            10
          ],
          "stride": [
            10,
            1
          ],
          "dtype": "torch.float32",
          "storage_offset": 0,
          "logical_bytes": 400000
        }
      ],
      "unique_storages": [
        {
          "storage_id": 0,
          "device": "cuda:0",
          "address": 46637089030144,
          "bytes": 6724,
          "aliases": [
            "parameters.weights.0"
          ]
        },
        {
          "storage_id": 1,
          "device": "cuda:0",
          "address": 46637089128448,
          "bytes": 6724,
          "aliases": [
            "gradients.weights.0"
          ]
        },
        {
          "storage_id": 2,
          "device": "cuda:0",
          "address": 46637089037312,
          "bytes": 6724,
          "aliases": [
            "parameters.weights.1"
          ]
        },
        {
          "storage_id": 3,
          "device": "cuda:0",
          "address": 46637089135616,
          "bytes": 6724,
          "aliases": [
            "gradients.weights.1"
          ]
        },
        {
          "storage_id": 4,
          "device": "cuda:0",
          "address": 46637089044480,
          "bytes": 6724,
          "aliases": [
            "parameters.weights.2"
          ]
        },
        {
          "storage_id": 5,
          "device": "cuda:0",
          "address": 46637089142784,
          "bytes": 6724,
          "aliases": [
            "gradients.weights.2"
          ]
        },
        {
          "storage_id": 6,
          "device": "cuda:0",
          "address": 46637089051648,
          "bytes": 6724,
          "aliases": [
            "parameters.weights.3"
          ]
        },
        {
          "storage_id": 7,
          "device": "cuda:0",
          "address": 46637089149952,
          "bytes": 6724,
          "aliases": [
            "gradients.weights.3"
          ]
        },
        {
          "storage_id": 8,
          "device": "cuda:0",
          "address": 46637089058816,
          "bytes": 3280,
          "aliases": [
            "parameters.head.weight"
          ]
        },
        {
          "storage_id": 9,
          "device": "cuda:0",
          "address": 46637089157120,
          "bytes": 3280,
          "aliases": [
            "gradients.head.weight"
          ]
        },
        {
          "storage_id": 10,
          "device": "cuda:0",
          "address": 46637089062400,
          "bytes": 40,
          "aliases": [
            "parameters.head.bias"
          ]
        },
        {
          "storage_id": 11,
          "device": "cuda:0",
          "address": 46637089160704,
          "bytes": 40,
          "aliases": [
            "gradients.head.bias"
          ]
        },
        {
          "storage_id": 12,
          "device": "cuda:0",
          "address": 46637089062912,
          "bytes": 6724,
          "aliases": [
            "initial.0"
          ]
        },
        {
          "storage_id": 13,
          "device": "cuda:0",
          "address": 46637089070080,
          "bytes": 6724,
          "aliases": [
            "initial.1"
          ]
        },
        {
          "storage_id": 14,
          "device": "cuda:0",
          "address": 46637089077248,
          "bytes": 6724,
          "aliases": [
            "initial.2"
          ]
        },
        {
          "storage_id": 15,
          "device": "cuda:0",
          "address": 46637089084416,
          "bytes": 6724,
          "aliases": [
            "initial.3"
          ]
        },
        {
          "storage_id": 16,
          "device": "cuda:0",
          "address": 46637089091584,
          "bytes": 3280,
          "aliases": [
            "initial.4"
          ]
        },
        {
          "storage_id": 17,
          "device": "cuda:0",
          "address": 46637089095168,
          "bytes": 40,
          "aliases": [
            "initial.5"
          ]
        },
        {
          "storage_id": 18,
          "device": "cuda:0",
          "address": 46637089095680,
          "bytes": 6724,
          "aliases": [
            "velocity.0"
          ]
        },
        {
          "storage_id": 19,
          "device": "cuda:0",
          "address": 46637089102848,
          "bytes": 6724,
          "aliases": [
            "velocity.1"
          ]
        },
        {
          "storage_id": 20,
          "device": "cuda:0",
          "address": 46637089110016,
          "bytes": 6724,
          "aliases": [
            "velocity.2"
          ]
        },
        {
          "storage_id": 21,
          "device": "cuda:0",
          "address": 46637089117184,
          "bytes": 6724,
          "aliases": [
            "velocity.3"
          ]
        },
        {
          "storage_id": 22,
          "device": "cuda:0",
          "address": 46637089124352,
          "bytes": 3280,
          "aliases": [
            "velocity.4"
          ]
        },
        {
          "storage_id": 23,
          "device": "cuda:0",
          "address": 46637089127936,
          "bytes": 40,
          "aliases": [
            "velocity.5"
          ]
        },
        {
          "storage_id": 24,
          "device": "cuda:0",
          "address": 46636734611456,
          "bytes": 3240000,
          "aliases": [
            "raw"
          ]
        },
        {
          "storage_id": 25,
          "device": "cuda:0",
          "address": 46636737851904,
          "bytes": 3240000,
          "aliases": [
            "raw_query"
          ]
        },
        {
          "storage_id": 26,
          "device": "cuda:0",
          "address": 46637089161216,
          "bytes": 80000,
          "aliases": [
            "labels"
          ]
        },
        {
          "storage_id": 27,
          "device": "cuda:0",
          "address": 46636741092352,
          "bytes": 3280000,
          "aliases": [
            "x"
          ]
        },
        {
          "storage_id": 28,
          "device": "cuda:0",
          "address": 46636744372736,
          "bytes": 3280000,
          "aliases": [
            "q"
          ]
        },
        {
          "storage_id": 29,
          "device": "cuda:0",
          "address": 46637089241600,
          "bytes": 160000,
          "aliases": [
            "orders"
          ]
        },
        {
          "storage_id": 30,
          "device": "cuda:0",
          "address": 46637089401856,
          "bytes": 80000,
          "aliases": [
            "order"
          ]
        },
        {
          "storage_id": 31,
          "device": "cuda:0",
          "address": 46637089482240,
          "bytes": 80000,
          "aliases": [
            "predictions"
          ]
        },
        {
          "storage_id": 32,
          "device": "cuda:0",
          "address": 46637089562624,
          "bytes": 400000,
          "aliases": [
            "scores"
          ]
        }
      ],
      "unique_named_storage_bytes": 13960864,
      "allocator_allocated_bytes": 81084416,
      "allocated_bytes_without_named_storage_reference": 67123552,
      "scope": "Actual CUDA storages reachable through named task tensors, model parameters/gradients, initial copies and momentum. Counts aliases once. Unnamed graph-pool and library allocations are not assigned to named buffers.",
      "is_per_kernel_working_set": false,
      "cache_residency_established": false
    },
    "cache_reference": {
      "bytes": 41943040,
      "unit": "40 MiB",
      "meaning": "A100 L2 nominal capacity reference, not measured cache residency or a per-kernel working set"
    },
    "profile": null,
    "energy": {
      "calibration_ms": 48.29452896118164,
      "repeats_per_trial": 104,
      "trials": [
        {
          "trial": 0,
          "invocations": 104,
          "idle_before": {
            "start": {
              "energy_mj": 13459591609,
              "time_s": 10.7185159845,
              "query_latency_s": 0.003647838999999209
            },
            "end": {
              "energy_mj": 13459769324,
              "time_s": 13.722142361,
              "query_latency_s": 0.003076716000000701
            },
            "duration_s": 3.0036263764999998,
            "energy_j": 177.715,
            "average_power_w": 59.16681295330875,
            "settle_seconds": 3
          },
          "idle_after": {
            "start": {
              "energy_mj": 13460331654,
              "time_s": 20.6704473955,
              "query_latency_s": 0.0027880890000027136
            },
            "end": {
              "energy_mj": 13460509800,
              "time_s": 23.6734891115,
              "query_latency_s": 0.0028400309999980777
            },
            "duration_s": 3.003041716000002,
            "energy_j": 178.146,
            "average_power_w": 59.32185325659987,
            "settle_seconds": 3
          },
          "active": {
            "start": {
              "energy_mj": 13459775468,
              "time_s": 13.7345870885,
              "query_latency_s": 0.002567564999999661
            },
            "end": {
              "energy_mj": 13460118608,
              "time_s": 17.666196035,
              "query_latency_s": 0.0034013440000002504
            },
            "duration_s": 3.931608946499999,
            "energy_j": 343.14,
            "average_power_w": 87.27724569491339
          },
          "telemetry_before": {
            "temperature_c": 37,
            "power_w": 60.941,
            "graphics_clock_mhz": 1095,
            "memory_clock_mhz": 1215
          },
          "telemetry_after": {
            "temperature_c": 39,
            "power_w": 87.97,
            "graphics_clock_mhz": 1410,
            "memory_clock_mhz": 1215
          },
          "paired_idle_w": 59.24433310495431,
          "idle_adjusted_mj_per_task": 1059.7543262993806,
          "unadjusted_mj_per_task": 3299.423076923077,
          "cuda_ms_per_task": 37.77359713040865,
          "wall_ms_per_task": 37.803932177884604
        },
        {
          "trial": 1,
          "invocations": 104,
          "idle_before": {
            "start": {
              "energy_mj": 13460693967,
              "time_s": 26.6775100285,
              "query_latency_s": 0.002935802999999737
            },
            "end": {
              "energy_mj": 13460878100,
              "time_s": 29.681656224999998,
              "query_latency_s": 0.0036562879999983977
            },
            "duration_s": 3.004146196499999,
            "energy_j": 184.133,
            "average_power_w": 61.292955787080345,
            "settle_seconds": 3
          },
          "idle_after": {
            "start": {
              "energy_mj": 13461434114,
              "time_s": 36.6538106095,
              "query_latency_s": 0.002946962999999414
            },
            "end": {
              "energy_mj": 13461618709,
              "time_s": 39.657499902,
              "query_latency_s": 0.002979244000002268
            },
            "duration_s": 3.003689292499999,
            "energy_j": 184.595,
            "average_power_w": 61.45609016915323,
            "settle_seconds": 3
          },
          "active": {
            "start": {
              "energy_mj": 13460878100,
              "time_s": 29.685530368499997,
              "query_latency_s": 0.003464844999999883
            },
            "end": {
              "energy_mj": 13461225481,
              "time_s": 33.649986346,
              "query_latency_s": 0.002896571999997377
            },
            "duration_s": 3.964455977500002,
            "energy_j": 347.381,
            "average_power_w": 87.62387625730668
          },
          "telemetry_before": {
            "temperature_c": 37,
            "power_w": 61.205,
            "graphics_clock_mhz": 1095,
            "memory_clock_mhz": 1215
          },
          "telemetry_after": {
            "temperature_c": 39,
            "power_w": 88.849,
            "graphics_clock_mhz": 1410,
            "memory_clock_mhz": 1215
          },
          "paired_idle_w": 61.37452297811679,
          "idle_adjusted_mj_per_task": 1000.6192837807083,
          "unadjusted_mj_per_task": 3340.201923076923,
          "cuda_ms_per_task": 38.087512676532455,
          "wall_ms_per_task": 38.1197690144231
        },
        {
          "trial": 2,
          "invocations": 104,
          "idle_before": {
            "start": {
              "energy_mj": 13461803087,
              "time_s": 42.6615290335,
              "query_latency_s": 0.0029097629999981223
            },
            "end": {
              "energy_mj": 13461987329,
              "time_s": 45.665077258500006,
              "query_latency_s": 0.0029115729999986684
            },
            "duration_s": 3.003548225000003,
            "energy_j": 184.242,
            "average_power_w": 61.341448912477446,
            "settle_seconds": 3
          },
          "idle_after": {
            "start": {
              "energy_mj": 13462543602,
              "time_s": 52.6148480635,
              "query_latency_s": 0.003108767000000512
            },
            "end": {
              "energy_mj": 13462728385,
              "time_s": 55.618927024499996,
              "query_latency_s": 0.002922302999998294
            },
            "duration_s": 3.0040789609999976,
            "energy_j": 184.783,
            "average_power_w": 61.51070008442436,
            "settle_seconds": 3
          },
          "active": {
            "start": {
              "energy_mj": 13461987329,
              "time_s": 45.668232741500006,
              "query_latency_s": 0.0029106429999998795
            },
            "end": {
              "energy_mj": 13462328117,
              "time_s": 49.610161099500004,
              "query_latency_s": 0.004133638999995526
            },
            "duration_s": 3.9419283579999984,
            "energy_j": 340.788,
            "average_power_w": 86.45210390705942
          },
          "telemetry_before": {
            "temperature_c": 37,
            "power_w": 61.205,
            "graphics_clock_mhz": 1095,
            "memory_clock_mhz": 1215
          },
          "telemetry_after": {
            "temperature_c": 39,
            "power_w": 87.97,
            "graphics_clock_mhz": 1410,
            "memory_clock_mhz": 1215
          },
          "paired_idle_w": 61.4260744984509,
          "idle_adjusted_mj_per_task": 948.5655289801525,
          "unadjusted_mj_per_task": 3276.8076923076924,
          "cuda_ms_per_task": 37.86791287935697,
          "wall_ms_per_task": 37.90315728846152
        }
      ],
      "summary": {
        "idle_adjusted_mj_per_task": {
          "mean": 1002.9797130200805,
          "sample_sd": 55.63196816552206,
          "values": [
            1059.7543262993806,
            1000.6192837807083,
            948.5655289801525
          ]
        },
        "unadjusted_mj_per_task": {
          "mean": 3305.477564102564,
          "sample_sd": 32.12786539688782,
          "values": [
            3299.423076923077,
            3340.201923076923,
            3276.8076923076924
          ]
        },
        "cuda_ms_per_task": {
          "mean": 37.90967422876603,
          "sample_sd": 0.1610706374635057,
          "values": [
            37.77359713040865,
            38.087512676532455,
            37.86791287935697
          ]
        },
        "wall_ms_per_task": {
          "mean": 37.94228616025641,
          "sample_sd": 0.1615132449200883,
          "values": [
            37.803932177884604,
            38.1197690144231,
            37.90315728846152
          ]
        }
      },
      "hardware": {
        "gpu": "NVIDIA A100-SXM4-40GB",
        "uuid": "GPU-059d2560-15ff-0871-a190-9065becce1b2",
        "total_memory_bytes": 42405855232,
        "power_limit_w": 400.0,
        "driver": "580.95.05",
        "cuda_pci_bus_id": "0000:15:00.0",
        "nvml_pci_bus_id": "00000000:15:00.0",
        "nvml_handle_selected_by_cuda_pci_bus_id": true
      },
      "software": {
        "python": "3.11.15",
        "torch": "2.5.1+cu124",
        "numpy": "2.2.6",
        "cuda": "12.4",
        "nvidia_ml_py": "12.560.30"
      },
      "repeat_predictions_equal": true,
      "repeat_state_hashes_equal": true,
      "state_hash_keys_checked": [
        "input_hashes",
        "epoch_permutation_sha256",
        "initial_parameter_sha256",
        "final_parameter_sha256",
        "final_velocity_sha256",
        "state_vector_sha256",
        "predictions_sha256",
        "scores_sha256"
      ],
      "scope": "Every task resets weights and momentum, normalizes raw9x9 inputs, trains from scratch, predicts allqueries; allinvocations included. GPUresident measurement excludes transfers, allocation, JIT/graphcapture, hostpreprocessing and verification.",
      "before_metadata": {
        "config": {
          "depth": 2,
          "alpha": 0.5,
          "seed": 11,
          "epochs": 2,
          "learning_rate": 0.1,
          "batch_size": 128,
          "momentum": 0.9,
          "weight_decay": 0.0
        },
        "parameter_count": 7554,
        "train_count": 10000,
        "test_count": 10000,
        "fresh_initialization_per_run": true,
        "checkpoint_loaded": false,
        "normalization": "4*x-0.5 then append one zero coordinate",
        "initialization": "branch normal(0,0.05/sqrt(41)); standard torch Linear head; seed-only",
        "optimizer": "SGD momentum; zero initial velocity; gradient mean cross entropy",
        "shuffle": "independent PCG64 permutation each epoch, seed-only",
        "augmentation": "none",
        "input_hashes": {
          "train_images": "aa68fb7d65b7dcd9c91fd62e48b42ed6a60f7e1fad7c53b1b0ec2c73a5a7c5d1",
          "train_labels": "72abc25a7a08e37d3ae171b30b3e5617ce288cafab5f4df112b352b93f2133fb",
          "test_images": "78aa1bca675869c1e08ba944b99d45dce473bee70d893240be8d733acc6d0144"
        },
        "epoch_permutation_sha256": [
          "06a0c75a115780a2112f0ba86be13d19162762e59c36bb0c27ae35dc688c01ef",
          "fc7dcea80db84b1177efbc923b2d1ecf4eb4184bd2d4afd32c9af9c6160502b0"
        ],
        "initial_parameter_sha256": {
          "weights.0": "194490b683f6426b65457a6725f0646768f8d46a361d9b125cfc768ab614f8fb",
          "weights.1": "bcbcbf732d68415f6c7c393c31343810467d8a86951e70237f6c36712785dd02",
          "weights.2": "60acda6b3f2f888c9f70f9c2d0ebb53e66d3deb3e1ba760160b87d76f0a92b98",
          "weights.3": "16f24aea0689177ce23b7d09de49339a0fa0a0ba75753a10a85ed850a7083cc8",
          "head.weight": "39bc6f6f4720618f6da778924cc306b6f5a0865d4042c58771d4c93cb573b512",
          "head.bias": "5ce9e9fee433d39fa1edf54bebdd7bf9e6686a86a38b65c7862577d8faf1379d"
        },
        "final_parameter_sha256": {
          "weights.0": "50d5bc3dcbe7dfd34cd1590048629b4d5ac4d4e7e3a0dac5dd5c3bd480e4dd2f",
          "weights.1": "fca2f7dc38383ba59ac71e98de5e71bd1fba71675ee0672f634a66bb8d153885",
          "weights.2": "e24b86ba7f45e01ebfc234ad3c15a876aa269ee0f286be37b2fc9bf0b386af75",
          "weights.3": "b170a498622c1698238ba8440f1cc25ae44d045a51bf0fbd39de495cfe73f99f",
          "head.weight": "706e7bcfd6b48a76582fd28f9679eadaf323fbb3c461557384f581dc2b1be842",
          "head.bias": "d98a0d189c7772e8352c850faa5f721eccff42bbc24dcce4c0776a2e83524312"
        },
        "final_velocity_sha256": {
          "weights.0": "450834df50ef4dfe42dcfcebe089aacea6efb8743f812892a20503ad20df3c84",
          "weights.1": "107a687d6251073024abbe9834c1e85b06eb5bb77480b1025f500d2e08e2c777",
          "weights.2": "b23132c42a1869bb803108a61ce23ea65fe8bcf98f4c6696fbdd72830237b6a4",
          "weights.3": "f38971086fd2090afea2f1e23deb312bc1af3cb41c7a0538c5b089882719e0c5",
          "head.weight": "715c14d2794695066fab32ab119b8e87599325fb7aba0a996f10b8b2c0a74c83",
          "head.bias": "fb1586654860b60c70e83fb404a3fecd2a495f4f841df3d7ccd9811e76aed328"
        },
        "state_vector_sha256": "fb2a11ad2a30522e01e9228619406c13c8f8d57ebcb75d10556f689fe4430ef9",
        "predictions_sha256": "37724115cb2d3242eb4a4f76d3a879ac92d729939320e5f6f2bd046bd631f200",
        "scores_sha256": "be5b7c1c7c0321aa6d157309cf6953c741bc4abfb768dfeea24d984f98ebf5e9",
        "preparation_seconds": 2.195325211,
        "run_count": 2,
        "validation": {
          "passed": true,
          "device": "cpu",
          "checks": [
            {
              "dtype": "torch.float64",
              "depth": 1,
              "passed": true,
              "forward_bitwise_equal": true,
              "maximum_gradient_abs_error": 1.3877787807814457e-17,
              "inverse_max_abs_error": 2.220446049250313e-16,
              "saved_activation_storages": 1,
              "saved_activation_bytes": 1968,
              "saved_parameter_references": 2
            },
            {
              "dtype": "torch.float64",
              "depth": 2,
              "passed": true,
              "forward_bitwise_equal": true,
              "maximum_gradient_abs_error": 1.3877787807814457e-17,
              "inverse_max_abs_error": 4.440892098500626e-16,
              "saved_activation_storages": 1,
              "saved_activation_bytes": 1968,
              "saved_parameter_references": 4
            },
            {
              "dtype": "torch.float32",
              "depth": 1,
              "passed": true,
              "forward_bitwise_equal": true,
              "maximum_gradient_abs_error": 0.0,
              "inverse_max_abs_error": 2.9802322387695312e-08,
              "saved_activation_storages": 1,
              "saved_activation_bytes": 984,
              "saved_parameter_references": 2
            },
            {
              "dtype": "torch.float32",
              "depth": 2,
              "passed": true,
              "forward_bitwise_equal": true,
              "maximum_gradient_abs_error": 1.4901161193847656e-08,
              "inverse_max_abs_error": 2.384185791015625e-07,
              "saved_activation_storages": 1,
              "saved_activation_bytes": 984,
              "saved_parameter_references": 4
            }
          ],
          "scope": "Identical weights; ordinary autograd versus custom reconstruction for inputs and every parameter. First-order differentiation only."
        },
        "reversible_core": true,
        "retained_core_states": 1,
        "core_saved_bytes_per_example": 328,
        "head_reversible": false,
        "input_lift_injective": true,
        "timing_scope": "run resets weights/gradients/momentum, normalizes raw train and query pixels, copies resident epoch permutations, trains and predicts. Excludes setup, CPU permutation generation, allocation, transfer, capture and diagnostics."
      },
      "after_metadata": {
        "config": {
          "depth": 2,
          "alpha": 0.5,
          "seed": 11,
          "epochs": 2,
          "learning_rate": 0.1,
          "batch_size": 128,
          "momentum": 0.9,
          "weight_decay": 0.0
        },
        "parameter_count": 7554,
        "train_count": 10000,
        "test_count": 10000,
        "fresh_initialization_per_run": true,
        "checkpoint_loaded": false,
        "normalization": "4*x-0.5 then append one zero coordinate",
        "initialization": "branch normal(0,0.05/sqrt(41)); standard torch Linear head; seed-only",
        "optimizer": "SGD momentum; zero initial velocity; gradient mean cross entropy",
        "shuffle": "independent PCG64 permutation each epoch, seed-only",
        "augmentation": "none",
        "input_hashes": {
          "train_images": "aa68fb7d65b7dcd9c91fd62e48b42ed6a60f7e1fad7c53b1b0ec2c73a5a7c5d1",
          "train_labels": "72abc25a7a08e37d3ae171b30b3e5617ce288cafab5f4df112b352b93f2133fb",
          "test_images": "78aa1bca675869c1e08ba944b99d45dce473bee70d893240be8d733acc6d0144"
        },
        "epoch_permutation_sha256": [
          "06a0c75a115780a2112f0ba86be13d19162762e59c36bb0c27ae35dc688c01ef",
          "fc7dcea80db84b1177efbc923b2d1ecf4eb4184bd2d4afd32c9af9c6160502b0"
        ],
        "initial_parameter_sha256": {
          "weights.0": "194490b683f6426b65457a6725f0646768f8d46a361d9b125cfc768ab614f8fb",
          "weights.1": "bcbcbf732d68415f6c7c393c31343810467d8a86951e70237f6c36712785dd02",
          "weights.2": "60acda6b3f2f888c9f70f9c2d0ebb53e66d3deb3e1ba760160b87d76f0a92b98",
          "weights.3": "16f24aea0689177ce23b7d09de49339a0fa0a0ba75753a10a85ed850a7083cc8",
          "head.weight": "39bc6f6f4720618f6da778924cc306b6f5a0865d4042c58771d4c93cb573b512",
          "head.bias": "5ce9e9fee433d39fa1edf54bebdd7bf9e6686a86a38b65c7862577d8faf1379d"
        },
        "final_parameter_sha256": {
          "weights.0": "50d5bc3dcbe7dfd34cd1590048629b4d5ac4d4e7e3a0dac5dd5c3bd480e4dd2f",
          "weights.1": "fca2f7dc38383ba59ac71e98de5e71bd1fba71675ee0672f634a66bb8d153885",
          "weights.2": "e24b86ba7f45e01ebfc234ad3c15a876aa269ee0f286be37b2fc9bf0b386af75",
          "weights.3": "b170a498622c1698238ba8440f1cc25ae44d045a51bf0fbd39de495cfe73f99f",
          "head.weight": "706e7bcfd6b48a76582fd28f9679eadaf323fbb3c461557384f581dc2b1be842",
          "head.bias": "d98a0d189c7772e8352c850faa5f721eccff42bbc24dcce4c0776a2e83524312"
        },
        "final_velocity_sha256": {
          "weights.0": "450834df50ef4dfe42dcfcebe089aacea6efb8743f812892a20503ad20df3c84",
          "weights.1": "107a687d6251073024abbe9834c1e85b06eb5bb77480b1025f500d2e08e2c777",
          "weights.2": "b23132c42a1869bb803108a61ce23ea65fe8bcf98f4c6696fbdd72830237b6a4",
          "weights.3": "f38971086fd2090afea2f1e23deb312bc1af3cb41c7a0538c5b089882719e0c5",
          "head.weight": "715c14d2794695066fab32ab119b8e87599325fb7aba0a996f10b8b2c0a74c83",
          "head.bias": "fb1586654860b60c70e83fb404a3fecd2a495f4f841df3d7ccd9811e76aed328"
        },
        "state_vector_sha256": "fb2a11ad2a30522e01e9228619406c13c8f8d57ebcb75d10556f689fe4430ef9",
        "predictions_sha256": "37724115cb2d3242eb4a4f76d3a879ac92d729939320e5f6f2bd046bd631f200",
        "scores_sha256": "be5b7c1c7c0321aa6d157309cf6953c741bc4abfb768dfeea24d984f98ebf5e9",
        "preparation_seconds": 2.195325211,
        "run_count": 315,
        "validation": {
          "passed": true,
          "device": "cpu",
          "checks": [
            {
              "dtype": "torch.float64",
              "depth": 1,
              "passed": true,
              "forward_bitwise_equal": true,
              "maximum_gradient_abs_error": 1.3877787807814457e-17,
              "inverse_max_abs_error": 2.220446049250313e-16,
              "saved_activation_storages": 1,
              "saved_activation_bytes": 1968,
              "saved_parameter_references": 2
            },
            {
              "dtype": "torch.float64",
              "depth": 2,
              "passed": true,
              "forward_bitwise_equal": true,
              "maximum_gradient_abs_error": 1.3877787807814457e-17,
              "inverse_max_abs_error": 4.440892098500626e-16,
              "saved_activation_storages": 1,
              "saved_activation_bytes": 1968,
              "saved_parameter_references": 4
            },
            {
              "dtype": "torch.float32",
              "depth": 1,
              "passed": true,
              "forward_bitwise_equal": true,
              "maximum_gradient_abs_error": 0.0,
              "inverse_max_abs_error": 2.9802322387695312e-08,
              "saved_activation_storages": 1,
              "saved_activation_bytes": 984,
              "saved_parameter_references": 2
            },
            {
              "dtype": "torch.float32",
              "depth": 2,
              "passed": true,
              "forward_bitwise_equal": true,
              "maximum_gradient_abs_error": 1.4901161193847656e-08,
              "inverse_max_abs_error": 2.384185791015625e-07,
              "saved_activation_storages": 1,
              "saved_activation_bytes": 984,
              "saved_parameter_references": 4
            }
          ],
          "scope": "Identical weights; ordinary autograd versus custom reconstruction for inputs and every parameter. First-order differentiation only."
        },
        "reversible_core": true,
        "retained_core_states": 1,
        "core_saved_bytes_per_example": 328,
        "head_reversible": false,
        "input_lift_injective": true,
        "timing_scope": "run resets weights/gradients/momentum, normalizes raw train and query pixels, copies resident epoch permutations, trains and predicts. Excludes setup, CPU permutation generation, allocation, transfer, capture and diagnostics."
      }
    },
    "metadata": {
      "config": {
        "depth": 2,
        "alpha": 0.5,
        "seed": 11,
        "epochs": 2,
        "learning_rate": 0.1,
        "batch_size": 128,
        "momentum": 0.9,
        "weight_decay": 0.0
      },
      "parameter_count": 7554,
      "train_count": 10000,
      "test_count": 10000,
      "fresh_initialization_per_run": true,
      "checkpoint_loaded": false,
      "normalization": "4*x-0.5 then append one zero coordinate",
      "initialization": "branch normal(0,0.05/sqrt(41)); standard torch Linear head; seed-only",
      "optimizer": "SGD momentum; zero initial velocity; gradient mean cross entropy",
      "shuffle": "independent PCG64 permutation each epoch, seed-only",
      "augmentation": "none",
      "input_hashes": {
        "train_images": "aa68fb7d65b7dcd9c91fd62e48b42ed6a60f7e1fad7c53b1b0ec2c73a5a7c5d1",
        "train_labels": "72abc25a7a08e37d3ae171b30b3e5617ce288cafab5f4df112b352b93f2133fb",
        "test_images": "78aa1bca675869c1e08ba944b99d45dce473bee70d893240be8d733acc6d0144"
      },
      "epoch_permutation_sha256": [
        "06a0c75a115780a2112f0ba86be13d19162762e59c36bb0c27ae35dc688c01ef",
        "fc7dcea80db84b1177efbc923b2d1ecf4eb4184bd2d4afd32c9af9c6160502b0"
      ],
      "initial_parameter_sha256": {
        "weights.0": "194490b683f6426b65457a6725f0646768f8d46a361d9b125cfc768ab614f8fb",
        "weights.1": "bcbcbf732d68415f6c7c393c31343810467d8a86951e70237f6c36712785dd02",
        "weights.2": "60acda6b3f2f888c9f70f9c2d0ebb53e66d3deb3e1ba760160b87d76f0a92b98",
        "weights.3": "16f24aea0689177ce23b7d09de49339a0fa0a0ba75753a10a85ed850a7083cc8",
        "head.weight": "39bc6f6f4720618f6da778924cc306b6f5a0865d4042c58771d4c93cb573b512",
        "head.bias": "5ce9e9fee433d39fa1edf54bebdd7bf9e6686a86a38b65c7862577d8faf1379d"
      },
      "final_parameter_sha256": {
        "weights.0": "50d5bc3dcbe7dfd34cd1590048629b4d5ac4d4e7e3a0dac5dd5c3bd480e4dd2f",
        "weights.1": "fca2f7dc38383ba59ac71e98de5e71bd1fba71675ee0672f634a66bb8d153885",
        "weights.2": "e24b86ba7f45e01ebfc234ad3c15a876aa269ee0f286be37b2fc9bf0b386af75",
        "weights.3": "b170a498622c1698238ba8440f1cc25ae44d045a51bf0fbd39de495cfe73f99f",
        "head.weight": "706e7bcfd6b48a76582fd28f9679eadaf323fbb3c461557384f581dc2b1be842",
        "head.bias": "d98a0d189c7772e8352c850faa5f721eccff42bbc24dcce4c0776a2e83524312"
      },
      "final_velocity_sha256": {
        "weights.0": "450834df50ef4dfe42dcfcebe089aacea6efb8743f812892a20503ad20df3c84",
        "weights.1": "107a687d6251073024abbe9834c1e85b06eb5bb77480b1025f500d2e08e2c777",
        "weights.2": "b23132c42a1869bb803108a61ce23ea65fe8bcf98f4c6696fbdd72830237b6a4",
        "weights.3": "f38971086fd2090afea2f1e23deb312bc1af3cb41c7a0538c5b089882719e0c5",
        "head.weight": "715c14d2794695066fab32ab119b8e87599325fb7aba0a996f10b8b2c0a74c83",
        "head.bias": "fb1586654860b60c70e83fb404a3fecd2a495f4f841df3d7ccd9811e76aed328"
      },
      "state_vector_sha256": "fb2a11ad2a30522e01e9228619406c13c8f8d57ebcb75d10556f689fe4430ef9",
      "predictions_sha256": "37724115cb2d3242eb4a4f76d3a879ac92d729939320e5f6f2bd046bd631f200",
      "scores_sha256": "be5b7c1c7c0321aa6d157309cf6953c741bc4abfb768dfeea24d984f98ebf5e9",
      "preparation_seconds": 2.195325211,
      "run_count": 315,
      "validation": {
        "passed": true,
        "device": "cpu",
        "checks": [
          {
            "dtype": "torch.float64",
            "depth": 1,
            "passed": true,
            "forward_bitwise_equal": true,
            "maximum_gradient_abs_error": 1.3877787807814457e-17,
            "inverse_max_abs_error": 2.220446049250313e-16,
            "saved_activation_storages": 1,
            "saved_activation_bytes": 1968,
            "saved_parameter_references": 2
          },
          {
            "dtype": "torch.float64",
            "depth": 2,
            "passed": true,
            "forward_bitwise_equal": true,
            "maximum_gradient_abs_error": 1.3877787807814457e-17,
            "inverse_max_abs_error": 4.440892098500626e-16,
            "saved_activation_storages": 1,
            "saved_activation_bytes": 1968,
            "saved_parameter_references": 4
          },
          {
            "dtype": "torch.float32",
            "depth": 1,
            "passed": true,
            "forward_bitwise_equal": true,
            "maximum_gradient_abs_error": 0.0,
            "inverse_max_abs_error": 2.9802322387695312e-08,
            "saved_activation_storages": 1,
            "saved_activation_bytes": 984,
            "saved_parameter_references": 2
          },
          {
            "dtype": "torch.float32",
            "depth": 2,
            "passed": true,
            "forward_bitwise_equal": true,
            "maximum_gradient_abs_error": 1.4901161193847656e-08,
            "inverse_max_abs_error": 2.384185791015625e-07,
            "saved_activation_storages": 1,
            "saved_activation_bytes": 984,
            "saved_parameter_references": 4
          }
        ],
        "scope": "Identical weights; ordinary autograd versus custom reconstruction for inputs and every parameter. First-order differentiation only."
      },
      "reversible_core": true,
      "retained_core_states": 1,
      "core_saved_bytes_per_example": 328,
      "head_reversible": false,
      "input_lift_injective": true,
      "timing_scope": "run resets weights/gradients/momentum, normalizes raw train and query pixels, copies resident epoch permutations, trains and predicts. Excludes setup, CPU permutation generation, allocation, transfer, capture and diagnostics."
    },
    "same_process_reference_hashes_matched": [
      "input_hashes",
      "initial_parameter_sha256",
      "epoch_permutation_sha256",
      "final_parameter_sha256",
      "final_velocity_sha256",
      "state_vector_sha256",
      "predictions_sha256",
      "scores_sha256"
    ],
    "frozen_qualification_comparison": "Must be performed independently by the orchestrator",
    "predictions_file": "result.predictions.npy",
    "predictions_file_sha256": "fce8a9458d2158537c7d6207fb0c8e9cfcbba007e7a748976bf3e7cd002729f3",
    "test_labels_read": false
  }
}