{
  "protocol": "where_frozen_v1",
  "definition": "Ten fixed Where factors; pretrained predictions and explicitly specified experimental partitions, not ground-truth segmentation.",
  "methods": {
    "position": {
      "implementation": "retained_original",
      "method": "seeded PCG64 random connected radial blob",
      "area_fraction": 0.25,
      "regions": [
        "random"
      ]
    },
    "eccentricity": {
      "implementation": "retained_original",
      "method": "elliptically normalized image-center radius",
      "thresholds": [
        0.4,
        0.8
      ],
      "regions": [
        "center",
        "middle",
        "periphery"
      ]
    },
    "thing_stuff": {
      "implementation": "OneFormer panoptic inference",
      "model": "shi-labs/oneformer_ade20k_swin_large",
      "revision": "4a5bac8e64f82681a12db2e151a4c2f4ce6092b2",
      "definition": "official ADE20K isthing IDs; void belongs to neither; tables/cabinets may be things",
      "regions": [
        "thing",
        "stuff"
      ]
    },
    "background": {
      "implementation": "OneFormer semantic inference",
      "model": "shi-labs/oneformer_ade20k_swin_large",
      "revision": "4a5bac8e64f82681a12db2e151a4c2f4ce6092b2",
      "classes": [
        "wall",
        "ceiling",
        "windowpane",
        "door",
        "curtain",
        "blind",
        "building",
        "house",
        "skyscraper",
        "sky",
        "mountain",
        "hill"
      ],
      "definition": "benchmark backdrop mapping over official ADE20K semantic labels; exclude table, desk, counter and floor; not a universal binary background label",
      "regions": [
        "background",
        "non_background"
      ]
    },
    "boundary": {
      "implementation": "retained_original",
      "method": "inner erosion of original Mask2Former semantic + GroundingDINO/SAM instance regions",
      "erosion_iterations": "max(1, round(2 * min(H,W)/256))",
      "regions": [
        "boundary",
        "interior"
      ]
    },
    "embodiment": {
      "implementation": "RobotSeg independent one-frame inference",
      "model": "showlab/RobotSeg",
      "revision": "dafb8c0d507276e2f96d2b07ac3661a7b3a41a5f",
      "category": "robot",
      "threshold": "mask logits > 0",
      "distance_band": "max(1, round(24 * min(H,W)/256))",
      "regions": [
        "embodiment",
        "near_embodiment",
        "environment"
      ]
    },
    "focal": {
      "implementation": "Sa2VA-8B referring segmentation",
      "model": "ByteDance/Sa2VA-8B",
      "revision": "902b049acb12eb68454abf4e51ab1e23d1bc9efa",
      "input": "full episode instruction, never current subtask",
      "definition": "visible directly manipulated object(s)/part(s); exclude robot and destinations/supports unless manipulated",
      "distance_band": "max(1, round(24 * min(H,W)/256))",
      "empty_prediction": "explicit missing entry; no fallback",
      "regions": [
        "focal",
        "near_focal",
        "far_focal"
      ]
    },
    "density": {
      "implementation": "retained_original",
      "method": "spatial_density_objects",
      "formula": "D(x)=sum_j exp(-distance(x, M_j)^2/(2*sigma^2))",
      "sigma": "32 * min(H,W)/256",
      "instances": "original grounded object masks; robot excluded; require at least two objects",
      "cuts": [
        0.3333333333333333,
        0.6666666666666666
      ],
      "regions": [
        "sparse",
        "intermediate",
        "dense"
      ]
    },
    "saliency": {
      "implementation": "DeepGaze IIE",
      "model": "matthias-k/DeepGaze",
      "revision": "c7db17e2d1d7ea6468ffdee2cfaddf141095dcff",
      "definition": "image-conditioned visual fixation log density with uniform center-bias input",
      "cuts": [
        0.3333333333333333,
        0.6666666666666666
      ],
      "regions": [
        "low",
        "medium",
        "high"
      ]
    },
    "depth": {
      "implementation": "retained_original",
      "model": "depth-anything/Depth-Anything-V2-Small-hf",
      "definition": "relative inverse depth, larger means nearer",
      "cuts": [
        0.3333333333333333,
        0.6666666666666666
      ],
      "regions": [
        "far",
        "mid",
        "near"
      ]
    }
  },
  "composition": {
    "baseline": "GaussianBlur(I, sigma=16.64*min(H,W)/256)",
    "alpha_sigma": "max(1, min(H,W)/256)",
    "alpha_kernel": "2*ceil(3*alpha_sigma)+1",
    "alpha": "Gaussian feathered binary mask, rounded to uint8",
    "removal": "(1-alpha)*I + alpha*baseline",
    "retain": "alpha*I + (1-alpha)*baseline",
    "note": "alpha in equations is saved uint8 / 255; exact integer blending; no inpainting"
  },
  "invalid_policy": "Preserve old skips; retain and flag new empty/full region masks; no invented masks. Missing concepts and degenerate regions are counted in validation.json.",
  "preserved_concepts": [
    "position",
    "eccentricity",
    "boundary",
    "density",
    "depth"
  ]
}