{
  "schema": "trading-ui-clef-aggregate-v1",
  "evaluation_date_jst": "2026-10-14",
  "environment": "synthetic mock order forms",
  "observation": "browser DOM and original image",
  "models": {
    "bf16-native-browser": {
      "action_exact_match": {
        "correct": 337,
        "total": 337,
        "rate": 1.0
      }
    },
    "packed-http-browser": {
      "action_exact_match": {
        "correct": 337,
        "total": 337,
        "rate": 1.0
      }
    }
  },
  "cohort": {
    "valid_episodes": 24,
    "invalid_or_incomplete_episodes": 6,
    "total_episodes": 30,
    "patterns": [
      "normal_limit",
      "normal_market",
      "normal_stop",
      "oco",
      "ifd",
      "ifdoco"
    ]
  },
  "packed_nvfp4_rollout": {
    "completed_handoff": {
      "correct": 24,
      "total": 24,
      "rate": 1.0
    },
    "final_order_all_fields_exact": {
      "correct": 24,
      "total": 24,
      "rate": 1.0
    },
    "negative_no_action": {
      "correct": 6,
      "total": 6,
      "rate": 1.0
    },
    "safety": {
      "episodes_with_complete_evidence": 30,
      "safe_episodes": {
        "correct": 30,
        "total": 30,
        "rate": 1.0
      },
      "unsafe_attempts": 0,
      "unsafe_effects": 0,
      "premature_handoff_attempts": 0,
      "interpretation": "blocked unsafe attempts and applied unsafe effects are reported separately; missing telemetry is not safe"
    }
  },
  "warm_serial_http_latency": {
    "requests": 336,
    "p50_seconds": 0.9780177030002051,
    "p95_seconds_nearest_rank": 1.2271018559999902,
    "min_seconds": 0.7288350479998371,
    "max_seconds": 1.471318599999904
  },
  "comparison": "semantic action comparison; HTTP numeric choice IDs differ from native option encoding",
  "limits": [
    "Policy-generated candidate lists include the correct action.",
    "Order and negative templates overlap across splits.",
    "This is not native AX training or UTM ten-case acceptance.",
    "No live broker operation or order submission.",
    "No BF16 serving speed comparison.",
    "Checkpoints within an order are correlated; aggregate counts do not establish performance on unseen screens."
  ],
  "source_report_sha256": "edad2598b97e44c9c37ffaa9954b4b67e1e6674a5813ab06d5932628c63d2e4c",
  "trained_checkpoint_sha256": "8190644ef9bcba5d48fe4cc555b61905ce1ae24d2f42710ee036915ace65419a",
  "dataset_sha256": "1bafec3a902b70cfe7258aaadecab50d48777df42e7499cd53e072d440403df7",
  "training_stage_comparison": {
    "dataset_sha256": "1bafec3a902b70cfe7258aaadecab50d48777df42e7499cd53e072d440403df7",
    "split": "test",
    "label_source": "teacher",
    "stages": [
      {
        "stage": "before_fine_tuning",
        "precision": "BF16",
        "measurement_status": "no_quantitative_score",
        "qualitative_observation": {
          "source": "user_report",
          "description": "DSH session logs showed unsuccessful behavior, including runs stopping partway through. No quantitative score was recorded."
        },
        "action_exact_match": null,
        "final_order_all_fields_exact": null,
        "completed_handoff": null
      },
      {
        "stage": "first_training_run",
        "run_id": "sbi-form-v3-lora-001",
        "precision": "BF16",
        "input_profile": "record",
        "action_exact_match": {
          "correct": 322,
          "total": 337,
          "rate": 0.9554896142433235
        },
        "final_order_all_fields_exact": {
          "correct": 16,
          "total": 24,
          "rate": 0.6666666666666666
        },
        "completed_handoff": {
          "correct": 16,
          "total": 24,
          "rate": 0.6666666666666666
        },
        "source_report": "sbi-three-model-test-v3-001/report/report.md",
        "source_report_sha256": "356e7544a38422a568234db64617b5c30873d107e858ab3b67d31aab6bc8b884"
      },
      {
        "stage": "second_training_run",
        "run_id": "sbi-form-browser-computer-50-50-v3-001",
        "precision": "BF16",
        "input_profile": "browser",
        "action_exact_match": {
          "correct": 337,
          "total": 337,
          "rate": 1.0
        },
        "final_order_all_fields_exact": {
          "correct": 24,
          "total": 24,
          "rate": 1.0
        },
        "completed_handoff": {
          "correct": 24,
          "total": 24,
          "rate": 1.0
        },
        "source_report": "sbi-browser-computer-50-50-v3-001/report/report.md",
        "source_report_sha256": "b6dca30bd909bacca1ec465d2322351252499d0718c03b88ffe27058202eb4f9"
      },
      {
        "stage": "quantized_second_training_run",
        "precision": "NVFP4",
        "input_profile": "browser",
        "action_exact_match": {
          "correct": 337,
          "total": 337,
          "rate": 1.0
        },
        "final_order_all_fields_exact": {
          "correct": 24,
          "total": 24,
          "rate": 1.0
        },
        "completed_handoff": {
          "correct": 24,
          "total": 24,
          "rate": 1.0
        },
        "source_report_sha256": "edad2598b97e44c9c37ffaa9954b4b67e1e6674a5813ab06d5932628c63d2e4c"
      }
    ],
    "comparison_note": "The first run uses the original record input profile. The second run and quantized result use the browser input profile; both training and input representation changed."
  },
  "model_weight_size_comparison": {
    "measurement": "tensor_payload_bytes",
    "included_components": [
      "model_weights",
      "joint_decision_head"
    ],
    "gb_definition_bytes": 1000000000,
    "bf16": {
      "model_tensor_bytes": 54713457120,
      "decision_head_tensor_bytes": 256112648,
      "total_bytes": 54969569768
    },
    "nvfp4_mixed_precision": {
      "model_tensor_bytes": 22899664480,
      "decision_head_tensor_bytes": 256112648,
      "total_bytes": 23155777128
    },
    "bytes_saved": 31813792640,
    "remaining_fraction": 0.42124719596186305,
    "reduction_fraction": 0.5787528040381369,
    "source_export_manifest_sha256": "8190644ef9bcba5d48fe4cc555b61905ce1ae24d2f42710ee036915ace65419a",
    "source_bf16_index_sha256": "0424520f63f04816eee4d229a4487dc9744866ceb46513e591dd987e401b1663",
    "source_packed_index_sha256": "5f7cbcfcd17a52cfa00a3c1409b410a8acb4e7fd0a32cadeb00fbd11e6a34092",
    "source_decision_head_file_sha256": "a010ac04f078e699988e4049cbea5e62c962393f59fec366640b64e8d69a4953",
    "method": "Sum tensor element counts multiplied by stored dtype widths from the saved export manifest, plus decision-head tensor data offsets. File headers, auxiliary files, and runtime memory are outside this measurement."
  }
}