{
  "source": {
    "shared_windows": "results/episode_task_suite/shared_windows.npz",
    "windows_csv": "results/episode_task_suite/windows.csv",
    "feature_manifest": "results/episode_task_suite/feature_manifest.json"
  },
  "dataset_scope": {
    "sample_episode_count": 1,
    "num_windows": 1161,
    "feature_dim": 8378,
    "first_start_frame": 0,
    "last_end_frame": 5819,
    "warning": "Single public sample episode; these extension probes validate task design and pipeline mechanics, not cross-episode generalization."
  },
  "baselines": {
    "minimal": "Ridge classifiers/regressors/projections plus cosine retrieval on the committed feature tensor.",
    "neural_mlp": "Small one-hidden-layer PyTorch MLP heads using the same inputs, targets, chronological split, and evaluator."
  },
  "run_config": {
    "train_fraction": 0.7,
    "ridge_l2": 10.0,
    "seed": 7,
    "future_windows": 4,
    "neural_epochs": 25,
    "neural_hidden_dim": 128,
    "neural_batch_size": 128,
    "skip_neural": false
  },
  "task_specs": {
    "body_motion_intensity": {
      "direction": "A",
      "direction_name": "Human Modeling & Motion Understanding",
      "name": "Body and Hand Motion Intensity",
      "family": "classification",
      "case_study": "A window with a fast reach or pour should be classified as high motion; a steady holding window should be low motion.",
      "input": "Current non-mocap feature blocks: video, depth, camera pose/rotation, IMU, SLAM, calibration, and language context.",
      "middle_process": "Compute the target from hand/body joint changes between neighboring windows, hide the mocap blocks from the input, then classify high versus low motion using the train-set median as the threshold.",
      "output": "Binary label: high_motion or low_motion.",
      "minimal_baseline": "Ridge classifier on standardized non-mocap features.",
      "neural_baseline": "One-hidden-layer MLP binary classifier on the same input features.",
      "metric_name": "macro-F1",
      "metric_key": "macro_f1",
      "metric_direction": "higher",
      "current_limit": "This is a motion-energy proxy, not a SMPL/MANO body model or a generative motion prior."
    },
    "multi_view_consistency_retrieval": {
      "direction": "B",
      "direction_name": "3D/4D Reconstruction & Neural Rendering",
      "name": "Multi-View Consistency Retrieval",
      "family": "retrieval",
      "case_study": "Given the fisheye camera features for a pouring moment, retrieve the synchronized stereo-left view from the same time window.",
      "input": "Query side: fisheye_cam0 video feature block. Candidate side: stereo_left video feature block from held-out windows.",
      "middle_process": "Learn a projection from one camera-view feature space into another, then rank held-out candidate windows by cosine similarity.",
      "output": "Ranked candidate windows; the correct synchronized view should rank near the top.",
      "minimal_baseline": "Ridge projection followed by cosine nearest-neighbor retrieval.",
      "neural_baseline": "One-hidden-layer MLP projection followed by the same cosine retrieval evaluator.",
      "metric_name": "MRR",
      "metric_key": "mrr",
      "metric_direction": "higher",
      "current_limit": "This checks calibrated multi-view signal, but it is still feature retrieval, not NeRF, Gaussian Splatting, or novel-view synthesis."
    },
    "action_phase_progress": {
      "direction": "C",
      "direction_name": "Egocentric Vision & Interaction",
      "name": "Action Phase Progress Estimation",
      "family": "regression",
      "case_study": "Inside a Pour coffee action segment, estimate whether the current window is near the beginning, middle, or end of that action.",
      "input": "Current non-caption multimodal feature vector, so the label text cannot be copied directly from the language block.",
      "middle_process": "Convert contiguous action-label runs into a normalized 0-to-1 progress target, train on earlier windows, and regress progress for later windows.",
      "output": "A scalar progress value between 0.0 and 1.0 for the current action segment.",
      "minimal_baseline": "Ridge regressor on standardized non-caption features.",
      "neural_baseline": "One-hidden-layer MLP regressor on the same input features.",
      "metric_name": "MAE",
      "metric_key": "mae",
      "metric_direction": "lower",
      "current_limit": "This is an action-structure probe inside one episode, not a general intent model across homes, people, or tasks."
    },
    "ego_motion_forecast": {
      "direction": "D",
      "direction_name": "Scene Reconstruction & World Modeling",
      "name": "Short-Horizon Ego-Motion Forecasting",
      "family": "forecast",
      "case_study": "From the current sensors, predict how the camera translation will change over the next 20 frames while the wearer moves through the scene.",
      "input": "Current multimodal features excluding the camera-translation block and caption text.",
      "middle_process": "Build a future target from camera-translation difference at a four-window horizon, then regress that future ego-motion delta from current sensors.",
      "output": "A future camera-translation delta vector.",
      "minimal_baseline": "Ridge regressor with a 20-frame forecast horizon.",
      "neural_baseline": "One-hidden-layer MLP regressor with the same horizon and split.",
      "metric_name": "MAE",
      "metric_key": "mae",
      "metric_direction": "lower",
      "current_limit": "This is a compact world-model proxy; it does not build a persistent map, scene graph, or object permanence model."
    }
  },
  "tasks": {
    "body_motion_intensity": {
      "train_windows": 812,
      "test_windows": 348,
      "target_threshold_train_median": 0.476467490196228,
      "input_dim": 6257,
      "target_source": "hand/body joint delta between neighboring windows",
      "minimal": {
        "accuracy": 0.7959770114942529,
        "macro_f1": 0.7860390200815733,
        "positive_rate_true": 0.35919540229885055,
        "positive_rate_pred": 0.42528735632183906,
        "num_test": 348
      },
      "neural_mlp": {
        "accuracy": 0.8189655172413793,
        "macro_f1": 0.8143276958646329,
        "positive_rate_true": 0.35919540229885055,
        "positive_rate_pred": 0.4827586206896552,
        "num_test": 348
      },
      "neural_training": {
        "available": true,
        "epochs": 25,
        "hidden_dim": 128,
        "loss_history": [
          0.3771491206338253,
          0.21354173564264928,
          0.13244220665816603,
          0.10503638759622433,
          0.07905124759145558,
          0.06667329128739868,
          0.05577798591458739,
          0.04417470416852406,
          0.03781222320836166,
          0.03380378236617948,
          0.02867619156507142,
          0.02776166619595342,
          0.022370315707412554,
          0.019363264677119373,
          0.016666794081076318,
          0.013987633805685296,
          0.012903780970043443,
          0.011304825730163035,
          0.009188220317801232,
          0.007710238994360704,
          0.006825127093502197,
          0.006784962428265751,
          0.005260340341830195,
          0.005132695187452098,
          0.00538189195812189
        ]
      }
    },
    "multi_view_consistency_retrieval": {
      "train_windows": 813,
      "test_windows": 348,
      "query_block": "video_fisheye_cam0",
      "target_block": "video_stereo_left",
      "query_dim": 686,
      "target_dim": 686,
      "minimal": {
        "mrr": 0.552907407283783,
        "top1": 0.41954022988505746,
        "top5": 0.7068965517241379,
        "top10": 0.8304597701149425,
        "median_rank": 2.0,
        "num_test": 348
      },
      "neural_mlp": {
        "mrr": 0.3451290726661682,
        "top1": 0.22988505747126436,
        "top5": 0.4540229885057471,
        "top10": 0.5862068965517241,
        "median_rank": 7.0,
        "num_test": 348
      },
      "neural_training": {
        "available": true,
        "epochs": 25,
        "hidden_dim": 128,
        "loss_history": [
          0.9802072261532294,
          0.8302588426172367,
          0.7033629322843798,
          0.6110119350428658,
          0.5481111645551918,
          0.4936590702771261,
          0.45608002786853363,
          0.42634973314855373,
          0.39841723533720697,
          0.37698091557398344,
          0.3569651228108823,
          0.34343410762620297,
          0.330490266008424,
          0.31780097566700805,
          0.30736573823117097,
          0.29689836681695586,
          0.2889042267573569,
          0.28041428760115655,
          0.27232901398252884,
          0.26450822736064444,
          0.2603846002292223,
          0.25524227731708815,
          0.2491118810053681,
          0.24406919165584287,
          0.23761634876880552
        ]
      }
    },
    "action_phase_progress": {
      "train_windows": 813,
      "test_windows": 348,
      "input_dim": 7482,
      "target_source": "normalized position inside contiguous action-label runs",
      "minimal": {
        "mse": 0.18190430104732513,
        "mae": 0.3415461480617523,
        "r2": -1.1725970135340686,
        "num_test": 348
      },
      "neural_mlp": {
        "mse": 0.15401607751846313,
        "mae": 0.31263846158981323,
        "r2": -0.8395105143218342,
        "num_test": 348
      },
      "neural_training": {
        "available": true,
        "epochs": 25,
        "hidden_dim": 128,
        "loss_history": [
          3.4688629706700644,
          1.5451849149162038,
          1.326330089921001,
          0.8649262141477578,
          0.577952673880963,
          0.46515296056848315,
          0.4212383984199749,
          0.354089474949303,
          0.30210305217007427,
          0.23982326431614592,
          0.19178202004157016,
          0.19724293038487287,
          0.20587197774597346,
          0.18455821174903697,
          0.17869186793599476,
          0.15895180840407203,
          0.14212195294809518,
          0.1556165573103548,
          0.13423408677877097,
          0.11096753546599239,
          0.09754248720400155,
          0.07875784645074702,
          0.07185868289307915,
          0.0782155046860377,
          0.09717305779896979
        ]
      }
    },
    "ego_motion_forecast": {
      "train_windows": 810,
      "test_windows": 347,
      "forecast_horizon_windows": 4,
      "forecast_horizon_frames": 20,
      "input_dim": 7461,
      "target_dim": 21,
      "target_source": "future minus current camera_translation feature block",
      "minimal": {
        "mse": 4.988194465637207,
        "mae": 0.19888411462306976,
        "r2": -8529.573106085107,
        "num_test": 347
      },
      "neural_mlp": {
        "mse": 0.8564521074295044,
        "mae": 0.10089198499917984,
        "r2": -1463.6636605537994,
        "num_test": 347
      },
      "neural_training": {
        "available": true,
        "epochs": 25,
        "hidden_dim": 128,
        "loss_history": [
          1.03622907459,
          0.6008042420869992,
          0.4436667482794067,
          0.35586878966402125,
          0.27937634311340476,
          0.2187315085051972,
          0.18008612191971438,
          0.1596080208266223,
          0.14920606138529602,
          0.13056322325159003,
          0.12379058529565364,
          0.11550549081078282,
          0.11186948313757225,
          0.10269535618063844,
          0.09326774505553422,
          0.08536725770912053,
          0.08817933531087122,
          0.08303169007470579,
          0.0769105920177183,
          0.07583652247249344,
          0.07666844201308709,
          0.07737687951252785,
          0.07119835648271773,
          0.07334094380522951,
          0.07172362936867607
        ]
      }
    }
  }
}