Blind strong-agent semantic-mask input ablation

Three independent fresh agents; no agent could read cause labels, clean references, pseudo-GT, existing VLM decisions, or evaluation. No confidence threshold and no 15-frame cap: every transient + A/B/C decision becomes a raw SAM mask. The green pseudo-GT column is shown only after decisions and is a noisy RGB-difference diagnostic, not semantic ground truth.

15local4 masked frames

15video masked frames

15alltrain masked frames

23local4 selected masks

24video selected masks

24alltrain selected masks

Jump to frames where decisions or A/B/C selections differ: 0002 · 0004 · 0006 · 0012 · 0016 · 0024 · 0026 · 0028 · 0030

Decision summaries

{
  "variants": {
    "local4": {
      "decision_counts": {
        "persistent": 58,
        "transient": 23,
        "artifact": 7
      },
      "selected_proposals": 23,
      "masked_frames": 15,
      "mean_mask_fraction": 0.06111737731232565
    },
    "video": {
      "decision_counts": {
        "persistent": 58,
        "transient": 24,
        "artifact": 6
      },
      "selected_proposals": 24,
      "masked_frames": 15,
      "mean_mask_fraction": 0.06359085472283262
    },
    "alltrain": {
      "decision_counts": {
        "persistent": 56,
        "transient": 24,
        "artifact": 8
      },
      "selected_proposals": 24,
      "masked_frames": 15,
      "mean_mask_fraction": 0.06153390112259157
    }
  },
  "pairwise": {
    "local4_vs_video": {
      "pixel_iou": 0.9513901819156327,
      "frame_presence_agreement": 1.0
    },
    "local4_vs_alltrain": {
      "pixel_iou": 0.9831501904573917,
      "frame_presence_agreement": 1.0
    },
    "video_vs_alltrain": {
      "pixel_iou": 0.9583137168916174,
      "frame_presence_agreement": 1.0
    }
  },
  "proposal_count": 88,
  "disagreement_frames": [
    "0002",
    "0004",
    "0006",
    "0012",
    "0016",
    "0024",
    "0026",
    "0028",
    "0030"
  ]
}

frame 0001

proposal	local4	video	alltrain
proposal_01	persistent · 0.96 · none floor chairs and nearby potted plant This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.94 · none fixed floor chair and cushion Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none left floor chair and cushion The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_02	persistent · 0.97 · none large static floral arrangements and display umbrellas This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none central blossom arrangement Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none central flower arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.96 · none tall potted plant against the shoji window This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none right-side potted plant Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none right potted red-leaf arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

Input evidence used by the three conditions

local4

proposal	local4	video	alltrain
proposal_01	transient · 0.99 · A large paper umbrella held very close to the camera The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A large foreground folding fan Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A large foreground paper parasol The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	transient · 0.91 · A small fragmented region of the same foreground umbrella canopy The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.91 · A small segment on the foreground fan Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.72 · A blurred near-field portion of the foreground parasol The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_03	transient · 0.93 · A foreground umbrella shaft/rib segment The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.96 · C foreground fan ribs and canopy Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.78 · A foreground parasol shaft and edge The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.

proposal	local4	video	alltrain
proposal_01	persistent · 0.97 · none static floral display beside the umbrellas This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none left blossom arrangement Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none central flower arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_02	persistent · 0.96 · none tall potted plant on the cabinet This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none right potted flower arrangement Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none right potted red-leaf arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.95 · none cabinet-top ornaments and animal figurine This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.95 · none cabinet-top ornaments Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none cabinet-top ornaments and shelf The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.98 · A freestanding dark wooden lattice partition in the foreground The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A large freestanding wooden screen Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A freestanding dark wooden lattice screen The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	persistent · 0.95 · none potted plant and shoji background This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none right potted flowers Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none right potted red-leaf arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.93 · none dark cabinet door panel This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	transient · 0.97 · A lower panel of freestanding wooden screen Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A right portion of the freestanding lattice screen The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.

proposal	local4	video	alltrain
proposal_01	persistent · 0.95 · none large shoji window surface This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none shoji window and wall surface Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none rectangular shoji window The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_02	artifact · 0.92 · none featureless tatami floor residual The proposal follows a textureless surface or reconstruction/appearance discrepancy rather than a bounded physical object.	artifact · 0.82 · none floor/cabinet boundary residual The proposal follows an incoherent edge or low-texture residual rather than a stable object silhouette, with no ordered-sequence motion signature.	artifact · 0.82 · none empty tatami floor residual The proposal follows residual spill, blur, shadow, or reconstruction mismatch without a distinct removable object, and the SAM choices mainly cover background surface.
proposal_03	persistent · 0.97 · none static flower arrangements near the umbrellas This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none blossoms and umbrellas Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none central flower arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.98 · C large gold folding screen placed in the foreground The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · B large orange folding room divider Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · C large gold folding room divider The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	transient · 0.89 · B lower/side portion of the foreground gold folding screen The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.94 · A divider fold and narrow panel Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A lower gap and panel portion of the folding divider The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_03	transient · 0.94 · A single gold folding-screen panel The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.96 · A central divider panel Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A right panel of the folding divider The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.

proposal	local4	video	alltrain
proposal_01	persistent · 0.97 · none static flower arrangements This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none blossom arrangement Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none central flower arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_02	persistent · 0.98 · none large red display umbrella This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.98 · none large red umbrella Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none red paper parasol The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.98 · none small purple display umbrella This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none small umbrella Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none small purple paper parasol The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.98 · A tall freestanding white-framed mirror The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A upright mirror or reflective frame Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A tall freestanding white mirror The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	persistent · 0.96 · none static floral arrangement This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none blossom arrangement Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none central flower arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.95 · none cabinet-top circular shelf and lucky-cat ornament This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.95 · none cabinet-top ornaments Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none cabinet-top circular shelf and lucky cat The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	persistent · 0.98 · none red and purple display umbrellas This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.98 · none paired umbrellas Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none red and purple paper parasols The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_02	persistent · 0.97 · none flower arrangement beside the table This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none central blossoms Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none central flower arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.97 · none flower arrangement in the large brown vase This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none right flower vase Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none rear flower arrangement in brown vase The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.98 · A loose square floor cushion in the middle foreground The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A scattered floor cushions Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A loose patterned floor cushion The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	transient · 0.96 · A upright patterned cushion in front of the table The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.98 · A upright foreground cushion Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A upright patterned floor chair/cushion The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_03	transient · 0.98 · A loose brown floor cushion at lower right The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A flat floor cushion Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A loose brown patterned floor cushion The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.

proposal	local4	video	alltrain
proposal_01	transient · 0.96 · C large motion-blurred sheet or fabric crossing the scene The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · B large moving drape or cloth Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.86 · C large sweeping translucent fabric or cloth The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	persistent · 0.94 · none cabinet-top display and drawers This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.95 · none cabinet-top ornaments Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none cabinet-top ornaments and shelf The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.96 · none color calibration chart leaning against cabinet This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none color chart Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none color calibration chart The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.99 · A black-and-white paper umbrella placed in the foreground The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A foreground black-and-white parasol Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A black-and-white foreground paper parasol The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	persistent · 0.97 · none static flower arrangement This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none central flower arrangement Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none central flower arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.95 · none cabinet-top display This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.95 · none cabinet-top ornaments Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none cabinet-top ornaments and shelf The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	artifact · 0.91 · none featureless tatami floor patch The proposal follows a textureless surface or reconstruction/appearance discrepancy rather than a bounded physical object.	artifact · 0.86 · none small floor-edge residual The proposal follows an incoherent edge or low-texture residual rather than a stable object silhouette, with no ordered-sequence motion signature.	artifact · 0.82 · none empty tatami floor residual The proposal follows residual spill, blur, shadow, or reconstruction mismatch without a distinct removable object, and the SAM choices mainly cover background surface.
proposal_02	persistent · 0.96 · none rectangular shoji window This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none shoji window and wall Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none rectangular shoji window and cabinet-top objects The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.98 · none red and purple display umbrellas This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.98 · none red umbrella Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none red and purple paper parasols The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.97 · B large patterned cushion or cloth held close to camera The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A large patterned foreground board or cloth Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A large foreground patterned cushion or padded board The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	transient · 0.94 · C upper diagonal portion of the same foreground cloth The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.98 · A upper portion of patterned foreground board Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A upper edge of the same foreground patterned cushion The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.

proposal	local4	video	alltrain
proposal_01	transient · 0.98 · A tall freestanding white-framed mirror The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A freestanding full-length mirror Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A tall freestanding white mirror The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	artifact · 0.91 · none featureless tatami floor residual The proposal follows a textureless surface or reconstruction/appearance discrepancy rather than a bounded physical object.	artifact · 0.80 · none plain floor residual patch The proposal follows an incoherent edge or low-texture residual rather than a stable object silhouette, with no ordered-sequence motion signature.	artifact · 0.82 · none empty tatami floor residual The proposal follows residual spill, blur, shadow, or reconstruction mismatch without a distinct removable object, and the SAM choices mainly cover background surface.
proposal_03	persistent · 0.97 · none static flower arrangement behind table This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none blossom arrangement Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none central flower arrangement The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.97 · A person holding a tall white-framed mirror The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A person holding an upright mirror Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A person holding a tall white mirror frame The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	persistent · 0.95 · none cabinet-top display and drawers This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.95 · none cabinet-top ornaments Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none cabinet-top ornaments and shelf The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.95 · none small patterned cushion beside table This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none floor cushion Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none brown patterned seat cushion The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.99 · A large dark wooden lattice partition in foreground The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A large freestanding slatted screen Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A freestanding dark wooden lattice screen The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	persistent · 0.96 · none color calibration chart beside low cabinet This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none color chart Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none color calibration chart The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.95 · none cabinet-top ornaments and drawers This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.95 · none cabinet-top ornaments Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none cabinet-top ornaments and shelf The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	persistent · 0.96 · none teddy bear, table, and adjacent flowers This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none teddy bear and table area Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none teddy bear, table, and central flowers The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_02	persistent · 0.96 · none table and display umbrellas This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.94 · none table and umbrella region Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none table edge and paper parasols The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.97 · none round shoji opening This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none round shoji window Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none circular shoji window The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.99 · A large dark cloth or cushion crossing the top foreground The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A dark overhead drape or cloth Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A large dark cloth crossing the upper foreground The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	persistent · 0.91 · none small shoji/window patch behind railing This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.94 · none round window edge and rail Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	artifact · 0.76 · none small circular-window residual below the cloth edge The proposal follows residual spill, blur, shadow, or reconstruction mismatch without a distinct removable object, and the SAM choices mainly cover background surface.

proposal	local4	video	alltrain
proposal_01	persistent · 0.96 · none wall around round shoji opening This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.95 · none teddy bear and table flowers Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none green wall around circular shoji window The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_02	persistent · 0.96 · none rectangular shoji window This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.94 · none table and umbrella region Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none rectangular shoji window The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.98 · none red and purple display umbrellas This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none round shoji window Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none red and purple paper parasols The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.99 · A headless mannequin wearing a blue shirt in foreground The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A person entering at right with extended arm Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A headless mannequin wearing a blue shirt The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	persistent · 0.95 · none small patterned cushion beside table This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none fixed floor cushion Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none green patterned seat cushion The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.90 · none partially occluded cushion edge This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	artifact · 0.82 · none tiny floor/cushion-edge residual The proposal follows an incoherent edge or low-texture residual rather than a stable object silhouette, with no ordered-sequence motion signature.	transient · 0.70 · A small visible base or foot of the foreground mannequin The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.

Blind strong-agent semantic-mask input ablation

Decision summaries

frame 0001

local4

video/ordered window

all-train gallery

frame 0002 · variants differ

local4

video/ordered window

all-train gallery

frame 0003

local4

video/ordered window

all-train gallery

frame 0004 · variants differ

local4

video/ordered window

all-train gallery

frame 0005

local4

video/ordered window

all-train gallery

frame 0006 · variants differ

local4

video/ordered window

all-train gallery

frame 0007

local4

video/ordered window

all-train gallery

frame 0008

local4

video/ordered window

all-train gallery

frame 0009

local4

video/ordered window

all-train gallery

frame 0010

local4

video/ordered window

all-train gallery

frame 0011

local4

video/ordered window

all-train gallery

frame 0012 · variants differ

local4

video/ordered window

all-train gallery

frame 0013

local4

video/ordered window

all-train gallery

frame 0014

local4

video/ordered window

all-train gallery

frame 0015

local4

video/ordered window

all-train gallery

frame 0016 · variants differ

local4

video/ordered window

all-train gallery

frame 0017

local4

video/ordered window

all-train gallery

frame 0018

local4

video/ordered window

all-train gallery

frame 0019

local4

video/ordered window

all-train gallery

frame 0020

local4

proposal	local4	video	alltrain
proposal_01	persistent · 0.96 · none teddy bear and flower arrangement near table This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.97 · none table flower arrangement Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none teddy bear, table, and central flowers The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_02	persistent · 0.98 · none red and purple display umbrellas This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.98 · none umbrellas Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none red and purple paper parasols The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	artifact · 0.90 · none featureless tatami floor patch The proposal follows a textureless surface or reconstruction/appearance discrepancy rather than a bounded physical object.	artifact · 0.83 · none floor patch beside fixed cushion The proposal follows an incoherent edge or low-texture residual rather than a stable object silhouette, with no ordered-sequence motion signature.	artifact · 0.82 · none empty tatami floor residual near the chair The proposal follows residual spill, blur, shadow, or reconstruction mismatch without a distinct removable object, and the SAM choices mainly cover background surface.

proposal	local4	video	alltrain
proposal_01	persistent · 0.96 · none display umbrellas and cabinet region This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.98 · none umbrellas Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none right cabinet and red parasol The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_02	persistent · 0.96 · none rectangular shoji window This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none shoji window Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none rectangular shoji window The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	persistent · 0.95 · none black floor chair and adjacent cushion This scene element is consistent with the Round-0 static render and is corroborated from the allowed context viewpoints as part of the staged room.	persistent · 0.96 · none floor chair and cushion Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none black floor chair and cushion The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.

proposal	local4	video	alltrain
proposal_01	transient · 0.98 · A large potted bouquet held close to the camera The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.99 · A foreground flower bouquet and vase Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	transient · 0.93 · A large blurred foreground flower vase and stems The target contains this specific foreground/introduced instance, the Round-0 render omits it, and the all-30 observation gallery does not show it persisting at the same 3D location.
proposal_02	artifact · 0.90 · none floor and static cushion residual The proposal follows a textureless surface or reconstruction/appearance discrepancy rather than a bounded physical object.	persistent · 0.94 · none fixed floor cushion Tracks the fixed room geometry/decor across the ordered neighboring observations and is represented at the corresponding location in Round-0 renders.	persistent · 0.89 · none green patterned seat cushion The same instance is visible repeatedly at the same fixed scene location across the all-30 high-resolution gallery; the target-render difference is reconstruction or appearance error.
proposal_03	transient · 0.91 · A blurred leaves/petals belonging to the foreground bouquet The target contains this movable foreground object/part where the Round-0 static render does not; the four allowed context views do not establish it as fixed scene geometry.	transient · 0.91 · A foreground bouquet petal against window Absent from the synchronized Round-0 render and does not remain at a fixed 3D scene location across the ordered target-centered sequence.	artifact · 0.77 · none soft cast-shadow or blur patch on the shoji wall The proposal follows residual spill, blur, shadow, or reconstruction mismatch without a distinct removable object, and the SAM choices mainly cover background surface.