{
  "candidates": {
    "smolvlm2_2_2b_nf4": {
      "exact_output_determinism": {
        "all_identical": true,
        "assessment_valid": true,
        "group_count": 12,
        "groups": [
          {
            "case_id": "airport-01",
            "identical": true,
            "repetitions": 3,
            "task": "classification",
            "unique_output_count": 1
          },
          {
            "case_id": "airport-01",
            "identical": true,
            "repetitions": 3,
            "task": "description",
            "unique_output_count": 1
          },
          {
            "case_id": "airport-01",
            "identical": true,
            "repetitions": 3,
            "task": "priority",
            "unique_output_count": 1
          },
          {
            "case_id": "beach-01",
            "identical": true,
            "repetitions": 3,
            "task": "classification",
            "unique_output_count": 1
          },
          {
            "case_id": "beach-01",
            "identical": true,
            "repetitions": 3,
            "task": "description",
            "unique_output_count": 1
          },
          {
            "case_id": "beach-01",
            "identical": true,
            "repetitions": 3,
            "task": "priority",
            "unique_output_count": 1
          },
          {
            "case_id": "dense_residential-01",
            "identical": true,
            "repetitions": 3,
            "task": "classification",
            "unique_output_count": 1
          },
          {
            "case_id": "dense_residential-01",
            "identical": true,
            "repetitions": 3,
            "task": "description",
            "unique_output_count": 1
          },
          {
            "case_id": "dense_residential-01",
            "identical": true,
            "repetitions": 3,
            "task": "priority",
            "unique_output_count": 1
          },
          {
            "case_id": "farmland-01",
            "identical": true,
            "repetitions": 3,
            "task": "classification",
            "unique_output_count": 1
          },
          {
            "case_id": "farmland-01",
            "identical": true,
            "repetitions": 3,
            "task": "description",
            "unique_output_count": 1
          },
          {
            "case_id": "farmland-01",
            "identical": true,
            "repetitions": 3,
            "task": "priority",
            "unique_output_count": 1
          }
        ],
        "identical_group_count": 12
      },
      "first_repetition": [
        {
          "case_id": "airport-01",
          "output": "AIRPORT.",
          "output_sha256": "0eda2867a65b385c213f4ec980461b556d8c7f4b77001dbbb41a7d1da1656215",
          "score": {
            "contract_compliant": false,
            "correct": true,
            "parsed": "airport"
          },
          "task": "classification"
        },
        {
          "case_id": "airport-01",
          "output": "HIGH.",
          "output_sha256": "ce67038593eae92f029e55b267f453c2635e5dd68d56a66cbf0be3885f93b17c",
          "score": {
            "contract_compliant": false,
            "correct": true,
            "parsed": "high"
          },
          "task": "priority"
        },
        {
          "case_id": "airport-01",
          "output": "The image is an aerial view of an airport.",
          "output_sha256": "8eab6986195363f9dde93068098ad5ca58c9b4fb7a7eaaf2266b53c442e44942",
          "score": {
            "visible_class_cue_present": true,
            "visible_class_cues": [
              "airport"
            ],
            "within_word_limit": true,
            "word_count": 9
          },
          "task": "description"
        },
        {
          "case_id": "beach-01",
          "output": "BEACH.",
          "output_sha256": "5df7db9d61344c75c5bbfee4cad4cc6ac6ed8e9377b53fd3a15b80a77a85b85f",
          "score": {
            "contract_compliant": false,
            "correct": true,
            "parsed": "beach"
          },
          "task": "classification"
        },
        {
          "case_id": "beach-01",
          "output": "HIGH.",
          "output_sha256": "ce67038593eae92f029e55b267f453c2635e5dd68d56a66cbf0be3885f93b17c",
          "score": {
            "contract_compliant": false,
            "correct": false,
            "parsed": "high"
          },
          "task": "priority"
        },
        {
          "case_id": "beach-01",
          "output": "The image is an aerial view of a beach with sand and water.",
          "output_sha256": "b4c789fcb71d1c33e96929d947304cb07653136ecdf5969e71607d19c73b2a28",
          "score": {
            "visible_class_cue_present": true,
            "visible_class_cues": [
              "beach",
              "sand",
              "water"
            ],
            "within_word_limit": true,
            "word_count": 13
          },
          "task": "description"
        },
        {
          "case_id": "dense_residential-01",
          "output": "DENSE_RESIDENTIAL.",
          "output_sha256": "3869bd805ee52fbc3ff126a4c23310b96ac9eca6c6e34eb57d33e184c50de55f",
          "score": {
            "contract_compliant": false,
            "correct": true,
            "parsed": "dense_residential"
          },
          "task": "classification"
        },
        {
          "case_id": "dense_residential-01",
          "output": "HIGH.",
          "output_sha256": "ce67038593eae92f029e55b267f453c2635e5dd68d56a66cbf0be3885f93b17c",
          "score": {
            "contract_compliant": false,
            "correct": true,
            "parsed": "high"
          },
          "task": "priority"
        },
        {
          "case_id": "dense_residential-01",
          "output": "The image is an aerial view of a city with buildings and roads.",
          "output_sha256": "918e47e07af229ce3f48c396877c30b782bf81e623b2da3b1b1fe7ec93997e51",
          "score": {
            "visible_class_cue_present": true,
            "visible_class_cues": [
              "buildings",
              "roads"
            ],
            "within_word_limit": true,
            "word_count": 13
          },
          "task": "description"
        },
        {
          "case_id": "farmland-01",
          "output": "FARMLAND.",
          "output_sha256": "886d27480d133da08212099a30fa4690c9c6be7719d23e75bada680fc024a6e8",
          "score": {
            "contract_compliant": false,
            "correct": true,
            "parsed": "farmland"
          },
          "task": "classification"
        },
        {
          "case_id": "farmland-01",
          "output": "HIGH.",
          "output_sha256": "ce67038593eae92f029e55b267f453c2635e5dd68d56a66cbf0be3885f93b17c",
          "score": {
            "contract_compliant": false,
            "correct": false,
            "parsed": "high"
          },
          "task": "priority"
        },
        {
          "case_id": "farmland-01",
          "output": "The image is an aerial view of a field with different colored sections.",
          "output_sha256": "125a0cbda8c566096b090333c4610d6e674747a29d8ca948d7e4e6ce9a257aad",
          "score": {
            "visible_class_cue_present": true,
            "visible_class_cues": [
              "field"
            ],
            "within_word_limit": true,
            "word_count": 13
          },
          "task": "description"
        }
      ],
      "idle": {
        "mean_board_power_w": 4.068870813381576
      },
      "memory": {
        "maximum_observed_cgroup_current_bytes": 3834916864,
        "maximum_observed_cgroup_peak_bytes": null,
        "maximum_observed_cuda_peak_allocated_bytes": 1914322432,
        "maximum_observed_cuda_peak_reserved_bytes": 2359296000,
        "maximum_observed_process_high_water_bytes": 3765387264,
        "minimum_observed_host_available_bytes": 2942619648
      },
      "model": {
        "bytes": 8992155310,
        "id": "HuggingFaceTB/SmolVLM2-2.2B-Instruct",
        "license": "apache-2.0",
        "revision": "482adb537c021c86670beed01cd58990d01e72e4",
        "root": "/srv/nomos/models/smolvlm2-2.2b-instruct-482adb5"
      },
      "model_load": {
        "board_energy_j": 506.2645048031892,
        "cuda_allocated_bytes_after": 1466724864,
        "cuda_reserved_bytes_after": 2357198848,
        "duration_s": 127.218008793,
        "host_mem_available_bytes_after": 3741851648,
        "mean_board_power_w": 3.9795034492871717
      },
      "overall": {
        "board_energy_j": 2007.6901552173929,
        "duration_s": 183.684810509,
        "duration_weighted_board_power_w": 10.930082621714778,
        "maximum_junction_temperature_c": 59.875
      },
      "physical_feasibility_passed": true,
      "summary_evidence": {
        "completed.json": {
          "bytes": 1252,
          "path": "/srv/nomos/results/vlm-edge-smolvlm2-2.2b-nf4-15w-v1/completed.json",
          "sha256": "b7d87451537d1d865a4f3fde77f98c7119118f421d17211b5598e97400d3274a"
        },
        "idle-baseline.json": {
          "bytes": 41883,
          "path": "/srv/nomos/results/vlm-edge-smolvlm2-2.2b-nf4-15w-v1/idle-baseline.json",
          "sha256": "1633fa71734bf3a5b70bb2f788ca6fb42856bd579aa266acc541ad48694a1978"
        },
        "initial.json": {
          "bytes": 7880,
          "path": "/srv/nomos/results/vlm-edge-smolvlm2-2.2b-nf4-15w-v1/initial.json",
          "sha256": "bdf8fcaaa5de5ffbe24ca2bc983ab1c0bbe0dfb39619441e2316b041d4f07bb2"
        },
        "model-load.json": {
          "bytes": 2085,
          "path": "/srv/nomos/results/vlm-edge-smolvlm2-2.2b-nf4-15w-v1/model-load.json",
          "sha256": "5dadba9c9e36bd33d5ae5cf53b699162cd36ccc7a170b9ea8250b2e43eb0cb2d"
        },
        "trials.jsonl": {
          "bytes": 1133204,
          "path": "/srv/nomos/results/vlm-edge-smolvlm2-2.2b-nf4-15w-v1/trials.jsonl",
          "sha256": "64538eaa18b39fe68faff84e4e3151aea9b9ca17745ddafe1a41544452e9ee03"
        }
      },
      "tasks": {
        "classification": {
          "board_energy_j": {
            "count": 12,
            "max": 56.02519492978064,
            "mean": 53.76807336162326,
            "median": 53.53875267214194,
            "min": 51.89599713786527,
            "p95": 56.02519492978064
          },
          "contract_compliant": {
            "count": 0,
            "fraction": 0.0,
            "total": 12
          },
          "correct": {
            "count": 12,
            "fraction": 1.0,
            "total": 12
          },
          "idle_adjusted_energy_j": {
            "count": 12,
            "max": 35.264013074423985,
            "mean": 33.82141242710793,
            "median": 33.630923416851005,
            "min": 32.67840552248684,
            "p95": 35.264013074423985
          },
          "latency_s": {
            "count": 12,
            "max": 5.135969304,
            "mean": 4.902259582416667,
            "median": 4.8927159815,
            "min": 4.716962816,
            "p95": 5.135969304
          },
          "mean_board_power_w": {
            "count": 12,
            "max": 11.050242026543504,
            "mean": 10.968853403095338,
            "median": 10.964608110200295,
            "min": 10.897335059111905,
            "p95": 11.050242026543504
          },
          "output_tokens": {
            "count": 12,
            "max": 8,
            "mean": 5.75,
            "median": 5.5,
            "min": 4,
            "p95": 8
          },
          "telemetry_error_sample_count": 0,
          "time_to_first_token_s": {
            "count": 12,
            "max": 4.056074911,
            "mean": 4.032155289416667,
            "median": 4.0295266304999995,
            "min": 4.011809907,
            "p95": 4.056074911
          },
          "trial_count": 12
        },
        "description": {
          "board_energy_j": {
            "count": 12,
            "max": 62.58506589134261,
            "mean": 61.30364776792433,
            "median": 62.09110779802376,
            "min": 58.467425930577924,
            "p95": 62.58506589134261
          },
          "idle_adjusted_energy_j": {
            "count": 12,
            "max": 39.12903227456157,
            "mean": 38.26051707764525,
            "median": 38.71658348208183,
            "min": 36.5289424315888,
            "p95": 39.12903227456157
          },
          "latency_s": {
            "count": 12,
            "max": 5.766731703,
            "mean": 5.663274099166666,
            "median": 5.746915298999999,
            "min": 5.3788211,
            "p95": 5.766731703
          },
          "mean_board_power_w": {
            "count": 12,
            "max": 10.891188112504155,
            "mean": 10.825311232229218,
            "median": 10.828440168435844,
            "min": 10.757363641707327,
            "p95": 10.891188112504155
          },
          "output_tokens": {
            "count": 12,
            "max": 15,
            "mean": 14.0,
            "median": 15.0,
            "min": 11,
            "p95": 15
          },
          "telemetry_error_sample_count": 0,
          "time_to_first_token_s": {
            "count": 12,
            "max": 4.059193039,
            "mean": 4.0244061916666665,
            "median": 4.0246627485,
            "min": 4.009284777,
            "p95": 4.059193039
          },
          "trial_count": 12,
          "visible_class_cue_present": {
            "count": 12,
            "fraction": 1.0,
            "total": 12
          },
          "within_word_limit": {
            "count": 12,
            "fraction": 1.0,
            "total": 12
          }
        },
        "priority": {
          "board_energy_j": {
            "count": 12,
            "max": 52.543721876972626,
            "mean": 52.235791805235145,
            "median": 52.198121453441146,
            "min": 52.06078160969103,
            "p95": 52.543721876972626
          },
          "contract_compliant": {
            "count": 0,
            "fraction": 0.0,
            "total": 12
          },
          "correct": {
            "count": 6,
            "fraction": 0.5,
            "total": 12
          },
          "idle_adjusted_energy_j": {
            "count": 12,
            "max": 33.25416826528405,
            "mean": 32.94310306822994,
            "median": 32.86909199280298,
            "min": 32.75094698625685,
            "p95": 33.25416826528405
          },
          "latency_s": {
            "count": 12,
            "max": 4.771203019,
            "mean": 4.741533860833333,
            "median": 4.7394393489999995,
            "min": 4.71125487,
            "p95": 4.771203019
          },
          "mean_board_power_w": {
            "count": 12,
            "max": 11.092093394917196,
            "mean": 11.016786037642737,
            "median": 11.005513699375431,
            "min": 10.933166223959418,
            "p95": 11.092093394917196
          },
          "output_tokens": {
            "count": 12,
            "max": 4,
            "mean": 4.0,
            "median": 4.0,
            "min": 4,
            "p95": 4
          },
          "telemetry_error_sample_count": 0,
          "time_to_first_token_s": {
            "count": 12,
            "max": 4.064872286,
            "mean": 4.0378537586666665,
            "median": 4.0314377925,
            "min": 4.018841872,
            "p95": 4.064872286
          },
          "trial_count": 12
        }
      },
      "trial_count": 36,
      "weight_mode": "nf4"
    },
    "smolvlm2_500m_bf16": {
      "exact_output_determinism": {
        "all_identical": true,
        "assessment_valid": true,
        "group_count": 12,
        "groups": [
          {
            "case_id": "airport-01",
            "identical": true,
            "repetitions": 3,
            "task": "classification",
            "unique_output_count": 1
          },
          {
            "case_id": "airport-01",
            "identical": true,
            "repetitions": 3,
            "task": "description",
            "unique_output_count": 1
          },
          {
            "case_id": "airport-01",
            "identical": true,
            "repetitions": 3,
            "task": "priority",
            "unique_output_count": 1
          },
          {
            "case_id": "beach-01",
            "identical": true,
            "repetitions": 3,
            "task": "classification",
            "unique_output_count": 1
          },
          {
            "case_id": "beach-01",
            "identical": true,
            "repetitions": 3,
            "task": "description",
            "unique_output_count": 1
          },
          {
            "case_id": "beach-01",
            "identical": true,
            "repetitions": 3,
            "task": "priority",
            "unique_output_count": 1
          },
          {
            "case_id": "dense_residential-01",
            "identical": true,
            "repetitions": 3,
            "task": "classification",
            "unique_output_count": 1
          },
          {
            "case_id": "dense_residential-01",
            "identical": true,
            "repetitions": 3,
            "task": "description",
            "unique_output_count": 1
          },
          {
            "case_id": "dense_residential-01",
            "identical": true,
            "repetitions": 3,
            "task": "priority",
            "unique_output_count": 1
          },
          {
            "case_id": "farmland-01",
            "identical": true,
            "repetitions": 3,
            "task": "classification",
            "unique_output_count": 1
          },
          {
            "case_id": "farmland-01",
            "identical": true,
            "repetitions": 3,
            "task": "description",
            "unique_output_count": 1
          },
          {
            "case_id": "farmland-01",
            "identical": true,
            "repetitions": 3,
            "task": "priority",
            "unique_output_count": 1
          }
        ],
        "identical_group_count": 12
      },
      "first_repetition": [
        {
          "case_id": "airport-01",
          "output": "AIRPORT",
          "output_sha256": "d306bba875f72c283f7506cdf21f7b469e78424e67e6007bb1b7d51a70f516bd",
          "score": {
            "contract_compliant": true,
            "correct": true,
            "parsed": "airport"
          },
          "task": "classification"
        },
        {
          "case_id": "airport-01",
          "output": "AIRPORT or",
          "output_sha256": "30f958e9e4bb40fdd6a778834e96c1c1784da950a186301c4770d27a5de60f44",
          "score": {
            "contract_compliant": false,
            "correct": false,
            "parsed": null
          },
          "task": "priority"
        },
        {
          "case_id": "airport-01",
          "output": "A large airport with a large terminal building.",
          "output_sha256": "b20639d4b01a81fe0352a48f7e7fb3699fb29885f1703ae57235a84c7486c978",
          "score": {
            "visible_class_cue_present": true,
            "visible_class_cues": [
              "airport"
            ],
            "within_word_limit": true,
            "word_count": 8
          },
          "task": "description"
        },
        {
          "case_id": "beach-01",
          "output": "AIRPORT",
          "output_sha256": "d306bba875f72c283f7506cdf21f7b469e78424e67e6007bb1b7d51a70f516bd",
          "score": {
            "contract_compliant": true,
            "correct": false,
            "parsed": "airport"
          },
          "task": "classification"
        },
        {
          "case_id": "beach-01",
          "output": "AIRPORT or",
          "output_sha256": "30f958e9e4bb40fdd6a778834e96c1c1784da950a186301c4770d27a5de60f44",
          "score": {
            "contract_compliant": false,
            "correct": false,
            "parsed": null
          },
          "task": "priority"
        },
        {
          "case_id": "beach-01",
          "output": "A beach with a sandy beach and a blue sea.",
          "output_sha256": "b2734fa5eedb73017797c65626a9f33ef040f3658656403873333da0175a6798",
          "score": {
            "visible_class_cue_present": true,
            "visible_class_cues": [
              "beach",
              "sea"
            ],
            "within_word_limit": true,
            "word_count": 10
          },
          "task": "description"
        },
        {
          "case_id": "dense_residential-01",
          "output": "FARMLAND",
          "output_sha256": "3797f3766de7f1dc54438b7f38384ec1519b655a4fa1ff77957065e72ad8c6e2",
          "score": {
            "contract_compliant": true,
            "correct": false,
            "parsed": "farmland"
          },
          "task": "classification"
        },
        {
          "case_id": "dense_residential-01",
          "output": "HIGH",
          "output_sha256": "b43e72eed0488d784f1e3092b48320e2355e9d6313f7ac401542018eec20d8a3",
          "score": {
            "contract_compliant": true,
            "correct": true,
            "parsed": "high"
          },
          "task": "priority"
        },
        {
          "case_id": "dense_residential-01",
          "output": "A city with many houses and trees.",
          "output_sha256": "cb1f571b7756aff7533eb856a6b7657730fc07c6ba4c9c0f9bbe95a519cc63de",
          "score": {
            "visible_class_cue_present": true,
            "visible_class_cues": [
              "houses"
            ],
            "within_word_limit": true,
            "word_count": 7
          },
          "task": "description"
        },
        {
          "case_id": "farmland-01",
          "output": "AIRPORT",
          "output_sha256": "d306bba875f72c283f7506cdf21f7b469e78424e67e6007bb1b7d51a70f516bd",
          "score": {
            "contract_compliant": true,
            "correct": false,
            "parsed": "airport"
          },
          "task": "classification"
        },
        {
          "case_id": "farmland-01",
          "output": "AIRPORT or",
          "output_sha256": "30f958e9e4bb40fdd6a778834e96c1c1784da950a186301c4770d27a5de60f44",
          "score": {
            "contract_compliant": false,
            "correct": false,
            "parsed": null
          },
          "task": "priority"
        },
        {
          "case_id": "farmland-01",
          "output": "A large, gray, rectangular building with a flat roof.",
          "output_sha256": "611d2a1d4c9cdac4ec184c95129c415837e551d8ebb8b02217305329979575e7",
          "score": {
            "visible_class_cue_present": false,
            "visible_class_cues": [],
            "within_word_limit": true,
            "word_count": 9
          },
          "task": "description"
        }
      ],
      "idle": {
        "mean_board_power_w": 3.8250636353990717
      },
      "memory": {
        "maximum_observed_cgroup_current_bytes": 2667094016,
        "maximum_observed_cgroup_peak_bytes": null,
        "maximum_observed_cuda_peak_allocated_bytes": 1495558144,
        "maximum_observed_cuda_peak_reserved_bytes": 1797259264,
        "maximum_observed_process_high_water_bytes": 3717455872,
        "minimum_observed_host_available_bytes": 3266658304
      },
      "model": {
        "bytes": 2034857549,
        "id": "HuggingFaceTB/SmolVLM2-500M-Video-Instruct",
        "license": "apache-2.0",
        "revision": "7b375e1b73b11138ff12fe22c8f2822d8fe03467",
        "root": "/srv/nomos/models/smolvlm2-500m-video-instruct-7b375e1-runtime-v2"
      },
      "model_load": {
        "board_energy_j": 23.081891864871707,
        "cuda_allocated_bytes_after": 1036428288,
        "cuda_reserved_bytes_after": 1061158912,
        "duration_s": 4.729910117,
        "host_mem_available_bytes_after": 4587433984,
        "mean_board_power_w": 4.8799853049875
      },
      "overall": {
        "board_energy_j": 717.8704572940417,
        "duration_s": 88.813613873,
        "duration_weighted_board_power_w": 8.082887588839347,
        "maximum_junction_temperature_c": 54.968
      },
      "physical_feasibility_passed": true,
      "summary_evidence": {
        "completed.json": {
          "bytes": 1247,
          "path": "/srv/nomos/results/vlm-edge-smolvlm2-500m-15w-v1/completed.json",
          "sha256": "df7fc284985af14684730a9ac25d5ac7f7258d1021b2636cc857cd21773b7207"
        },
        "idle-baseline.json": {
          "bytes": 41473,
          "path": "/srv/nomos/results/vlm-edge-smolvlm2-500m-15w-v1/idle-baseline.json",
          "sha256": "3e737748bccaa845ff2366b6678c242c3369e2f7150fcc0ac01bbf99d349724c"
        },
        "initial.json": {
          "bytes": 7877,
          "path": "/srv/nomos/results/vlm-edge-smolvlm2-500m-15w-v1/initial.json",
          "sha256": "486708abd61627a239504f0a1d41279ab7b9bf787a3db41dc53cfa69439904dd"
        },
        "model-load.json": {
          "bytes": 2020,
          "path": "/srv/nomos/results/vlm-edge-smolvlm2-500m-15w-v1/model-load.json",
          "sha256": "d43c94c200751e9dbfc39249c3c694e02be469cee753c0d7ed8580b1fa98e2c6"
        },
        "trials.jsonl": {
          "bytes": 599046,
          "path": "/srv/nomos/results/vlm-edge-smolvlm2-500m-15w-v1/trials.jsonl",
          "sha256": "218faded21f183042bd1daf106828ae1fac91abb617e9875fcd288ea17186f9d"
        }
      },
      "tasks": {
        "classification": {
          "board_energy_j": {
            "count": 12,
            "max": 19.124416524123294,
            "mean": 18.32601984933476,
            "median": 18.16185709415857,
            "min": 17.903086442924703,
            "p95": 19.124416524123294
          },
          "contract_compliant": {
            "count": 12,
            "fraction": 1.0,
            "total": 12
          },
          "correct": {
            "count": 3,
            "fraction": 0.25,
            "total": 12
          },
          "idle_adjusted_energy_j": {
            "count": 12,
            "max": 10.07997917220967,
            "mean": 9.658160189554124,
            "median": 9.571045276752,
            "min": 9.335213164693519,
            "p95": 10.07997917220967
          },
          "latency_s": {
            "count": 12,
            "max": 2.364519447,
            "mean": 2.2660694006666664,
            "median": 2.2381041575,
            "min": 2.19393228,
            "p95": 2.364519447
          },
          "mean_board_power_w": {
            "count": 12,
            "max": 8.160273043124514,
            "mean": 8.088301023197234,
            "median": 8.115273526245804,
            "min": 7.7859331748425,
            "p95": 8.160273043124514
          },
          "output_tokens": {
            "count": 12,
            "max": 5,
            "mean": 4.25,
            "median": 4.0,
            "min": 4,
            "p95": 5
          },
          "telemetry_error_sample_count": 0,
          "time_to_first_token_s": {
            "count": 12,
            "max": 1.400361222,
            "mean": 1.37117846875,
            "median": 1.369417447,
            "min": 1.357419268,
            "p95": 1.400361222
          },
          "trial_count": 12
        },
        "description": {
          "board_energy_j": {
            "count": 12,
            "max": 25.022860012422758,
            "mean": 23.594168982108922,
            "median": 23.715598276582696,
            "min": 22.063412019352803,
            "p95": 25.022860012422758
          },
          "idle_adjusted_energy_j": {
            "count": 12,
            "max": 13.056834587702214,
            "mean": 12.353175923833794,
            "median": 12.329695286596154,
            "min": 11.665541182999114,
            "p95": 13.056834587702214
          },
          "latency_s": {
            "count": 12,
            "max": 3.132797841,
            "mean": 2.938772823083333,
            "median": 2.966905701,
            "min": 2.714212694,
            "p95": 3.132797841
          },
          "mean_board_power_w": {
            "count": 12,
            "max": 8.132487860994031,
            "mean": 8.031103708931681,
            "median": 8.015951485347504,
            "min": 7.959982781114885,
            "p95": 8.132487860994031
          },
          "output_tokens": {
            "count": 12,
            "max": 13,
            "mean": 11.0,
            "median": 11.0,
            "min": 9,
            "p95": 13
          },
          "telemetry_error_sample_count": 0,
          "time_to_first_token_s": {
            "count": 12,
            "max": 1.374142664,
            "mean": 1.3572563098333335,
            "median": 1.3533475775000001,
            "min": 1.349966694,
            "p95": 1.374142664
          },
          "trial_count": 12,
          "visible_class_cue_present": {
            "count": 9,
            "fraction": 0.75,
            "total": 12
          },
          "within_word_limit": {
            "count": 12,
            "fraction": 1.0,
            "total": 12
          }
        },
        "priority": {
          "board_energy_j": {
            "count": 12,
            "max": 18.37682476676861,
            "mean": 17.902349276393124,
            "median": 18.0877411989144,
            "min": 17.189310374237007,
            "p95": 18.37682476676861
          },
          "contract_compliant": {
            "count": 3,
            "fraction": 0.25,
            "total": 12
          },
          "correct": {
            "count": 3,
            "fraction": 0.25,
            "total": 12
          },
          "idle_adjusted_energy_j": {
            "count": 12,
            "max": 9.700126557946083,
            "mean": 9.501391598283321,
            "median": 9.600569918222693,
            "min": 9.112440487140685,
            "p95": 9.700126557946083
          },
          "latency_s": {
            "count": 12,
            "max": 2.268380094,
            "mean": 2.1962922656666666,
            "median": 2.212046317,
            "min": 2.111564841,
            "p95": 2.268380094
          },
          "mean_board_power_w": {
            "count": 12,
            "max": 8.18163919781623,
            "mean": 8.151209388314038,
            "median": 8.150989412102561,
            "min": 8.101298726512548,
            "p95": 8.18163919781623
          },
          "output_tokens": {
            "count": 12,
            "max": 4,
            "mean": 3.75,
            "median": 4.0,
            "min": 3,
            "p95": 4
          },
          "telemetry_error_sample_count": 0,
          "time_to_first_token_s": {
            "count": 12,
            "max": 1.371483699,
            "mean": 1.363153756,
            "median": 1.362941318,
            "min": 1.351785172,
            "p95": 1.371483699
          },
          "trial_count": 12
        }
      },
      "trial_count": 36,
      "weight_mode": "bf16"
    }
  },
  "created_at_utc": "2026-08-26T02:25:25.534567+00:00",
  "failed_baseline": {
    "completed_trial_count": 0,
    "evidence": {
      "initial.json": {
        "bytes": 7711,
        "path": "/srv/nomos/results/vlm-edge-smolvlm2-15w-smoke-v2/initial.json",
        "sha256": "60543d244ba65ad02f16af9610e553934c5751911b8e21874123ae78c41e8fda"
      },
      "interrupted.json": {
        "bytes": 1507,
        "path": "/srv/nomos/results/vlm-edge-smolvlm2-15w-smoke-v2/interrupted.json",
        "sha256": "12e574ec74b1b6aafd66aacc79f2c96373b8b1697d179e99e09b29d07fa27906"
      },
      "model-load.json": {
        "bytes": 2042,
        "path": "/srv/nomos/results/vlm-edge-smolvlm2-15w-smoke-v2/model-load.json",
        "sha256": "c0cbbe6bc157a1a16d1a2c921623a763f23446310b8f86ec5d60c7db2e55850c"
      }
    },
    "hardware_reset_reason": "BCCPLEXWDT",
    "interpretation": "The idle baseline persisted and no trial did. The reset therefore occurred after model load and before the first request could be committed. The hardware reset reason records the watchdog source but does not by itself prove root cause.",
    "model": {
      "bytes": 8992155310,
      "id": "HuggingFaceTB/SmolVLM2-2.2B-Instruct",
      "license": "apache-2.0",
      "revision": "482adb537c021c86670beed01cd58990d01e72e4",
      "root": "/srv/nomos/models/smolvlm2-2.2b-instruct-482adb5"
    },
    "model_load": {
      "cuda_allocated_bytes_after": 4504592896,
      "cuda_reserved_bytes_after": 4735369216,
      "duration_s": 112.132225313,
      "host_mem_available_bytes_after": 395436032
    },
    "weight_mode": "bf16"
  },
  "interpretation_contract": [
    "Physical feasibility is separate from task correctness.",
    "Parsed correctness is separate from strict output-contract compliance.",
    "A hardware watchdog reset is reported as an observed event, not proof of a root cause.",
    "All measured requests are retained; no outlier is removed."
  ],
  "limitations": [
    "Four AID aerial images are a calibration set, not a representative accuracy benchmark.",
    "AID imagery is not raw spacecraft L0 data or operator telemetry.",
    "Only the stock 15 W mode was tested; a second power mode requires supervised host control.",
    "These SmolVLM2 results do not reproduce Gemma 3n model quality or the 64 GB AGX Orin platform.",
    "The development board, models, and containers are not flight qualified.",
    "Description cue coverage is a lexical check, not a human semantic-quality score.",
    "The AID mirror license is unknown, so no dataset image is redistributed."
  ],
  "plain_conclusion": "NF4 made the 2.2B model feasible and raised parsed scene classification from 25% to 100% on four frozen images, at higher latency, power, and energy. It still applied the explicit priority policy correctly on only half the scenes and never obeyed the strict label-only output contract. Use the model for perception; keep policy and schema enforcement deterministic.",
  "protocol": {
    "payload": {
      "decoding": {
        "do_sample": false,
        "repetitions": 3
      },
      "expected_trials_per_power_mode": 36,
      "presentation_case_count": 4,
      "scene_classes": [
        {
          "description_terms": [
            "airport",
            "runway",
            "aircraft",
            "airplane",
            "tarmac"
          ],
          "directory": "Airport",
          "label": "airport",
          "priority": "high"
        },
        {
          "description_terms": [
            "beach",
            "coast",
            "shore",
            "sand",
            "ocean",
            "sea",
            "water"
          ],
          "directory": "Beach",
          "label": "beach",
          "priority": "routine"
        },
        {
          "description_terms": [
            "building",
            "buildings",
            "houses",
            "residential",
            "roads",
            "rooftops",
            "urban"
          ],
          "directory": "DenseResidential",
          "label": "dense_residential",
          "priority": "high"
        },
        {
          "description_terms": [
            "farmland",
            "agricultural",
            "agriculture",
            "field",
            "fields",
            "crops",
            "plots",
            "terraces"
          ],
          "directory": "Farmland",
          "label": "farmland",
          "priority": "routine"
        }
      ],
      "selection_seed": "nomos-vlm-edge-2026-08-25",
      "tasks": [
        {
          "max_new_tokens": 8,
          "name": "classification",
          "prompt": "Classify this overhead image. Reply with exactly one label and no other text: AIRPORT, BEACH, DENSE_RESIDENTIAL, or FARMLAND."
        },
        {
          "max_new_tokens": 4,
          "name": "priority",
          "prompt": "Apply this mission policy to the overhead image: AIRPORT or DENSE_RESIDENTIAL is HIGH; BEACH or FARMLAND is ROUTINE. Reply with exactly HIGH or ROUTINE and no other text."
        },
        {
          "max_new_tokens": 42,
          "name": "description",
          "prompt": "Describe only directly visible features in one sentence of at most 30 words. Do not infer a location, event, intent, or condition that is not visibly supported."
        }
      ],
      "version": "nomos-vlm-edge-v1"
    },
    "sha256": "d27000e217657a98ab56e476381954c5c504a3cd30d5707faa22b155dff7c3e2"
  },
  "purpose": "physical VLM scale, quantization, latency, power, and behavior comparison",
  "schema": 1,
  "shared_conditions": {
    "attention": "sdpa",
    "board": {
      "boot_id": "ee77e059-3591-4594-9ac0-bb3df6dd2e85",
      "gpu_state_before": {
        "available_frequencies": "306000000 408000000 510000000 612000000 714000000 816000000 918000000 1020000000",
        "cur_freq": "306000000",
        "governor": "nvhost_podgov",
        "max_freq": "612000000",
        "min_freq": "306000000"
      },
      "machine": "aarch64",
      "model": null,
      "platform": "Linux-5.15.148-tegra-aarch64-with-glibc2.39",
      "reset_reason_before": "BCCPLEXWDT"
    },
    "compute_dtype": "bfloat16",
    "dataset": {
      "id": "blanchon/AID",
      "license": "unknown",
      "revision": "d9b5531544523f29d365a77a007b2ace3cea6882",
      "root": "/srv/nomos/datasets/aid-vlm-edge-d9b5531",
      "selection_seed": "nomos-vlm-edge-2026-08-25"
    },
    "decoding": "greedy",
    "power_mode": "15W",
    "presentation_cases": [
      {
        "bytes": 304502,
        "case_id": "airport-01",
        "label": "airport",
        "priority": "high",
        "repo_path": "data/Airport/airport_47.jpg",
        "sha256": "b7b85ca239ab179fb7fa5448717fa8c05c65f8bcd988942098750a943d6e7f8b"
      },
      {
        "bytes": 286849,
        "case_id": "beach-01",
        "label": "beach",
        "priority": "routine",
        "repo_path": "data/Beach/beach_290.jpg",
        "sha256": "a938af3fcfead6de5fc949aba16b0b911bd04764ad01ff47d14cd73e6a422b31"
      },
      {
        "bytes": 485453,
        "case_id": "dense_residential-01",
        "label": "dense_residential",
        "priority": "high",
        "repo_path": "data/DenseResidential/denseresidential_115.jpg",
        "sha256": "8bdc903ed22ff8bf303c5b73be66209cb58c237e3f0242f1eacff1e33527eb46"
      },
      {
        "bytes": 79441,
        "case_id": "farmland-01",
        "label": "farmland",
        "priority": "routine",
        "repo_path": "data/Farmland/farmland_69.jpg",
        "sha256": "ac5f68a32a112c65699dca3e3371af6849f27e1bcb8e2c317f898e23021fdd29"
      }
    ],
    "repetitions": 3
  },
  "source_summaries": {
    "large": {
      "bytes": 17088,
      "path": "/srv/nomos/results/vlm-edge-smolvlm2-2.2b-nf4-15w-summary-v1.json",
      "sha256": "1f6ca5add4a2552d811f2d2ed42a73f827b1d7b2584a424692b4e91340340d07"
    },
    "small": {
      "bytes": 17031,
      "path": "/srv/nomos/results/vlm-edge-smolvlm2-500m-15w-summary-v1.json",
      "sha256": "8a391308526b32ba5365d9db616ea64fd92be046bf135c78e9f9ac2ab91d3f01"
    }
  },
  "tradeoffs": {
    "classification_accuracy_percentage_point_change": 75.0,
    "classification_mean_energy_ratio": 2.933974414721321,
    "classification_mean_power_ratio": 1.3561381273566209,
    "classification_median_latency_ratio": 2.186098428486566,
    "priority_accuracy_percentage_point_change": 25.0
  },
  "writer": {
    "bytes": 11114,
    "path": "/workspace/experiments/vlm_edge/compare.py",
    "sha256": "33dc7b05e0fd656f9920b5a39beb083db70ea556bd0b682faa6ae618006d9af9"
  }
}