scenesmith-example-data_scene_064

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view. 2. `local_context` is a highlighted context view; subject is red and targets are green.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: interaction_clearance
Subject object: kitchen_dining_chair_0
Target objects: none

Verified asset facts:
- subject `kitchen_dining_chair_0`: category `chair`, affordances [sittable, supportable, pushable, liftable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`
- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `front: [sit]; top: [place_or_reach_objects]`
- low-confidence candidates: `none`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_dining_chair_0`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], affordances [sittable, supportable, pushable, liftable], task targets [desk, table]
- targets: none

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_dining_chair_0`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`
Target cues:
- none
Nearby same-room context objects that may affect access or clearance:
- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_dining_chair_1`: category `chair`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `medium`
- `kitchen_eucalyptus_sprig_0_f0_1`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `small`
- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_eucalyptus_sprig_0_f0_2`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `interaction_clearance` metric for `kitchen_dining_chair_0` with affordance `sittable`.
Relevant context: the surrounding scene.
Check whether the immediate sit/stand zone at the usable sitting side has enough free space for a normal sit/stand motion. Consider the combined effect of nearby furniture around the seat edge; multiple pieces that collectively pinch the sit/stand zone should count as blocked even if no single item fully closes it. Do not require 360-degree circulation around the furniture.
If you choose `degraded` or `fail`, name the blocker or the directly constrained approach or sit/stand zone. Consider the combined effect of nearby furniture only when the local renders clearly show that the front and side access are genuinely pinched or eliminated. A standard coffee table in front of a sofa or loveseat is normal unless the local evidence shows it closes off the usable seating edge. Do not infer exact centimeter clearance from auxiliary geometry or bounding-box numbers.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "interaction_clearance__kitchen_dining_chair_0",
  "metric": "interaction_clearance",
  "subject_id": "kitchen_dining_chair_0",
  "target_ids": [],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view. 2. `local_context` is a highlighted context view; subject is red and targets are green.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: interaction_clearance\nSubject object: kitchen_dining_chair_0\nTarget objects: none\n\nVerified asset facts:\n- subject `kitchen_dining_chair_0`: category `chair`, affordances [sittable, supportable, pushable, liftable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`\n- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `front: [sit]; top: [place_or_reach_objects]`\n- low-confidence candidates: `none`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_dining_chair_0`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], affordances [sittable, supportable, pushable, liftable], task targets [desk, table]\n- targets: none\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_dining_chair_0`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`\nTarget cues:\n- none\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_dining_chair_1`: category `chair`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `medium`\n- `kitchen_eucalyptus_sprig_0_f0_1`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `small`\n- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_eucalyptus_sprig_0_f0_2`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `interaction_clearance` metric for `kitchen_dining_chair_0` with affordance `sittable`.\nRelevant context: the surrounding scene.\nCheck whether the immediate sit/stand zone at the usable sitting side has enough free space for a normal sit/stand motion. Consider the combined effect of nearby furniture around the seat edge; multiple pieces that collectively pinch the sit/stand zone should count as blocked even if no single item fully closes it. Do not require 360-degree circulation around the furniture.\nIf you choose `degraded` or `fail`, name the blocker or the directly constrained approach or sit/stand zone. Consider the combined effect of nearby furniture only when the local renders clearly show that the front and side access are genuinely pinched or eliminated. A standard coffee table in front of a sofa or loveseat is normal unless the local evidence shows it closes off the usable seating edge. Do not infer exact centimeter clearance from auxiliary geometry or bounding-box numbers.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_dining_chair_0_local.png",
    "render_check_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_dining_chair_0_local_context.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: interaction_clearance
Subject object: kitchen_dining_chair_1
Target objects: none

Verified asset facts:
- subject `kitchen_dining_chair_1`: category `chair`, affordances [sittable, supportable, pushable, liftable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`
- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `front: [sit]; top: [place_or_reach_objects]`
- low-confidence candidates: `none`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_dining_chair_1`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], affordances [sittable, supportable, pushable, liftable], task targets [desk, table]
- targets: none

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_dining_chair_1`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`
Target cues:
- none
Nearby same-room context objects that may affect access or clearance:
- `kitchen_dining_chair_0`: category `chair`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `medium`
- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `small`
- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_eucalyptus_sprig_0_f0_0`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_eucalyptus_sprig_0_f0_2`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `interaction_clearance` metric for `kitchen_dining_chair_1` with affordance `sittable`.
Relevant context: the surrounding scene.
Check whether the immediate sit/stand zone at the usable sitting side has enough free space for a normal sit/stand motion. Consider the combined effect of nearby furniture around the seat edge; multiple pieces that collectively pinch the sit/stand zone should count as blocked even if no single item fully closes it. Do not require 360-degree circulation around the furniture.
If you choose `degraded` or `fail`, name the blocker or the directly constrained approach or sit/stand zone. Consider the combined effect of nearby furniture only when the local renders clearly show that the front and side access are genuinely pinched or eliminated. A standard coffee table in front of a sofa or loveseat is normal unless the local evidence shows it closes off the usable seating edge. Do not infer exact centimeter clearance from auxiliary geometry or bounding-box numbers.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "interaction_clearance__kitchen_dining_chair_1",
  "metric": "interaction_clearance",
  "subject_id": "kitchen_dining_chair_1",
  "target_ids": [],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: interaction_clearance\nSubject object: kitchen_dining_chair_1\nTarget objects: none\n\nVerified asset facts:\n- subject `kitchen_dining_chair_1`: category `chair`, affordances [sittable, supportable, pushable, liftable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`\n- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `front: [sit]; top: [place_or_reach_objects]`\n- low-confidence candidates: `none`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_dining_chair_1`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], affordances [sittable, supportable, pushable, liftable], task targets [desk, table]\n- targets: none\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_dining_chair_1`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`\nTarget cues:\n- none\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_dining_chair_0`: category `chair`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `medium`\n- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `small`\n- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_eucalyptus_sprig_0_f0_0`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_eucalyptus_sprig_0_f0_2`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `interaction_clearance` metric for `kitchen_dining_chair_1` with affordance `sittable`.\nRelevant context: the surrounding scene.\nCheck whether the immediate sit/stand zone at the usable sitting side has enough free space for a normal sit/stand motion. Consider the combined effect of nearby furniture around the seat edge; multiple pieces that collectively pinch the sit/stand zone should count as blocked even if no single item fully closes it. Do not require 360-degree circulation around the furniture.\nIf you choose `degraded` or `fail`, name the blocker or the directly constrained approach or sit/stand zone. Consider the combined effect of nearby furniture only when the local renders clearly show that the front and side access are genuinely pinched or eliminated. A standard coffee table in front of a sofa or loveseat is normal unless the local evidence shows it closes off the usable seating edge. Do not infer exact centimeter clearance from auxiliary geometry or bounding-box numbers.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_dining_chair_1_local.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: interaction_clearance
Subject object: kitchen_fridge_0
Target objects: none

Verified asset facts:
- subject `kitchen_fridge_0`: no offline asset YAML loaded; heuristic category `refrigerator`, affordances [openable, containable], usable face `front`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_fridge_0`: category `refrigerator`, group `appliance_storage`, keywords [refrigerator, fridge, cooling cabinet], affordances [openable, containable], task targets [wall, clear_space]
- targets: none

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_fridge_0`: category `refrigerator`, room `kitchen`, placement `floor`, footprint `large`, height `mid`, elevation `floor_level`
Target cues:
- none
Nearby same-room context objects that may affect access or clearance:
- `kitchen_semi_flush_ceiling_light_0`: category `pathway light`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`
- `kitchen_kitchen_prep_cart_0`: category `serving cart`, placement `floor`, proximity `mid`, bbox_relation `overlapping_or_touching`, footprint `medium`
- `kitchen_canister_jar_0`: category `jar`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`
- `kitchen_kitchen_towel_0`: category `towel`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`
- `kitchen_teaspoon_0`: category `teaspoon_teaspoon`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`
- `kitchen_chefs_knife_0_s1_1`: category `chefs_knife_s1`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `small`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `interaction_clearance` metric for `kitchen_fridge_0` with affordance `openable`.
Relevant context: the surrounding scene.
First infer the most likely operating side from door panels, handles, hinges, or the side that faces into the room. Then check only that side for a normal standing zone and immediate hand/door operating clearance. Do not assume a large hinged swing arc unless the render clearly shows hinged doors needing that arc. Do not require the entire front face to be empty; a nearby bench or seat off to one side is acceptable when the floor directly before the door/handle area remains usable. A narrow diagonal side gap or a squeeze-by path does not count as clear operating space.
If you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "interaction_clearance__kitchen_fridge_0",
  "metric": "interaction_clearance",
  "subject_id": "kitchen_fridge_0",
  "target_ids": [],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: interaction_clearance\nSubject object: kitchen_fridge_0\nTarget objects: none\n\nVerified asset facts:\n- subject `kitchen_fridge_0`: no offline asset YAML loaded; heuristic category `refrigerator`, affordances [openable, containable], usable face `front`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_fridge_0`: category `refrigerator`, group `appliance_storage`, keywords [refrigerator, fridge, cooling cabinet], affordances [openable, containable], task targets [wall, clear_space]\n- targets: none\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_fridge_0`: category `refrigerator`, room `kitchen`, placement `floor`, footprint `large`, height `mid`, elevation `floor_level`\nTarget cues:\n- none\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_semi_flush_ceiling_light_0`: category `pathway light`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`\n- `kitchen_kitchen_prep_cart_0`: category `serving cart`, placement `floor`, proximity `mid`, bbox_relation `overlapping_or_touching`, footprint `medium`\n- `kitchen_canister_jar_0`: category `jar`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`\n- `kitchen_kitchen_towel_0`: category `towel`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`\n- `kitchen_teaspoon_0`: category `teaspoon_teaspoon`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`\n- `kitchen_chefs_knife_0_s1_1`: category `chefs_knife_s1`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `small`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `interaction_clearance` metric for `kitchen_fridge_0` with affordance `openable`.\nRelevant context: the surrounding scene.\nFirst infer the most likely operating side from door panels, handles, hinges, or the side that faces into the room. Then check only that side for a normal standing zone and immediate hand/door operating clearance. Do not assume a large hinged swing arc unless the render clearly shows hinged doors needing that arc. Do not require the entire front face to be empty; a nearby bench or seat off to one side is acceptable when the floor directly before the door/handle area remains usable. A narrow diagonal side gap or a squeeze-by path does not count as clear operating space.\nIf you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_fridge_0_local_operating_zone.png",
    "render_check_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_fridge_0_local_operating_zone_context.png",
    "render_check_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_fridge_0_local_depth_probe.png",
    "render_check_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_fridge_0_local_plan_context.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: interaction_clearance
Subject object: kitchen_base_cabinet_run_0
Target objects: none

Verified asset facts:
- subject `kitchen_base_cabinet_run_0`: no offline asset YAML loaded; heuristic category `cabinet`, affordances [openable, containable, supportable], usable face `top`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_base_cabinet_run_0`: category `cabinet`, group `storage`, keywords [cabinet, storage cabinet, cupboard], affordances [openable, containable, supportable], task targets [wall, clear_space]
- targets: none

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_base_cabinet_run_0`: category `cabinet`, room `kitchen`, placement `floor`, footprint `large`, height `low`, elevation `floor_level`
Target cues:
- none
Nearby same-room context objects that may affect access or clearance:
- `kitchen_hand_soap_dispenser_0`: category `soap dispenser`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_kitchen_prep_cart_0`: category `serving cart`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `medium`
- `kitchen_dish_drying_mat_0`: category `place mat`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`
- `kitchen_dish_soap_pump_bottle_0`: category `bottle`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_kitchen_sponge_scrubber_0`: category `sponge_scrubber_sponge`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_electric_kettle_1`: category `kettle`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `interaction_clearance` metric for `kitchen_base_cabinet_run_0` with affordance `openable`.
Relevant context: the surrounding scene.
First infer the most likely operating side from door panels, handles, hinges, or the side that faces into the room. Then check only that side for a normal standing zone and immediate hand/door operating clearance. Do not assume a large hinged swing arc unless the render clearly shows hinged doors needing that arc. Do not require the entire front face to be empty; a nearby bench or seat off to one side is acceptable when the floor directly before the door/handle area remains usable. A narrow diagonal side gap or a squeeze-by path does not count as clear operating space.
If you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "interaction_clearance__kitchen_base_cabinet_run_0",
  "metric": "interaction_clearance",
  "subject_id": "kitchen_base_cabinet_run_0",
  "target_ids": [],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: interaction_clearance\nSubject object: kitchen_base_cabinet_run_0\nTarget objects: none\n\nVerified asset facts:\n- subject `kitchen_base_cabinet_run_0`: no offline asset YAML loaded; heuristic category `cabinet`, affordances [openable, containable, supportable], usable face `top`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_base_cabinet_run_0`: category `cabinet`, group `storage`, keywords [cabinet, storage cabinet, cupboard], affordances [openable, containable, supportable], task targets [wall, clear_space]\n- targets: none\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_base_cabinet_run_0`: category `cabinet`, room `kitchen`, placement `floor`, footprint `large`, height `low`, elevation `floor_level`\nTarget cues:\n- none\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_hand_soap_dispenser_0`: category `soap dispenser`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_kitchen_prep_cart_0`: category `serving cart`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `medium`\n- `kitchen_dish_drying_mat_0`: category `place mat`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`\n- `kitchen_dish_soap_pump_bottle_0`: category `bottle`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_kitchen_sponge_scrubber_0`: category `sponge_scrubber_sponge`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_electric_kettle_1`: category `kettle`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `interaction_clearance` metric for `kitchen_base_cabinet_run_0` with affordance `openable`.\nRelevant context: the surrounding scene.\nFirst infer the most likely operating side from door panels, handles, hinges, or the side that faces into the room. Then check only that side for a normal standing zone and immediate hand/door operating clearance. Do not assume a large hinged swing arc unless the render clearly shows hinged doors needing that arc. Do not require the entire front face to be empty; a nearby bench or seat off to one side is acceptable when the floor directly before the door/handle area remains usable. A narrow diagonal side gap or a squeeze-by path does not count as clear operating space.\nIf you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_base_cabinet_run_0_local_operating_zone.png",
    "render_check_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_base_cabinet_run_0_local_operating_zone_context.png",
    "render_check_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_base_cabinet_run_0_local_depth_probe.png",
    "render_check_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_base_cabinet_run_0_local_plan_context.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: interaction_clearance
Subject object: kitchen_pepper_shaker_0
Target objects: kitchen_base_cabinet_run_0

Verified asset facts:
- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, affordances [graspable, containable, openable], usable face `front`, benchmark relevance `functional`, confidence `0.58`, source `asset_annotation`
- interaction facts: placement class `unknown`, access type `none`, interaction height `none`, surface map `front: [grasp, access_contents, open]`
- low-confidence candidates: `none`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, group `unknown`, keywords [cocktail shaker], affordances [graspable, containable, openable], task targets [table, desk, nightstand, coffee_table, shelf, cabinet]
- target `kitchen_base_cabinet_run_0`: category `cabinet`, group `storage`, keywords [cabinet, storage cabinet, cupboard], task targets [wall, clear_space]

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, room `kitchen`, placement `floor`, footprint `tiny`, height `very_low`, elevation `high`
Target cues:
- target `kitchen_base_cabinet_run_0`: category `cabinet`, room `kitchen`, placement `floor`, footprint `large`, height `low`, elevation `floor_level`, subject_proximity `far`
Nearby same-room context objects that may affect access or clearance:
- `kitchen_salt_shaker_0`: category `cocktail shaker`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_print_3`: category `print_print`, placement `floor`, proximity `touching`, bbox_relation `very_close`, footprint `small`
- `kitchen_floating_wall_shelf_0`: category `wall shelf`, placement `floor`, proximity `touching`, bbox_relation `very_close`, footprint `medium`
- `kitchen_picture_frame_0`: category `picture frame`, placement `floor`, proximity `near`, bbox_relation `very_close`, footprint `small`
- `kitchen_coffee_mug_3`: category `mug`, placement `floor`, proximity `near`, bbox_relation `very_close`, footprint `tiny`
- `kitchen_potted_basil_herb_plant_0`: category `potted plant`, placement `floor`, proximity `near`, bbox_relation `nearby`, footprint `tiny`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `interaction_clearance` metric for `kitchen_pepper_shaker_0` with affordance `openable`.
Relevant context: kitchen_base_cabinet_run_0.
First infer the most likely operating side from door panels, handles, hinges, or the side that faces into the room. Then check only that side for a normal standing zone and immediate hand/door operating clearance. Do not assume a large hinged swing arc unless the render clearly shows hinged doors needing that arc. Do not require the entire front face to be empty; a nearby bench or seat off to one side is acceptable when the floor directly before the door/handle area remains usable. A narrow diagonal side gap or a squeeze-by path does not count as clear operating space.
If you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "interaction_clearance__kitchen_pepper_shaker_0__to__kitchen_base_cabinet_run_0",
  "metric": "interaction_clearance",
  "subject_id": "kitchen_pepper_shaker_0",
  "target_ids": [
    "kitchen_base_cabinet_run_0"
  ],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: interaction_clearance\nSubject object: kitchen_pepper_shaker_0\nTarget objects: kitchen_base_cabinet_run_0\n\nVerified asset facts:\n- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, affordances [graspable, containable, openable], usable face `front`, benchmark relevance `functional`, confidence `0.58`, source `asset_annotation`\n- interaction facts: placement class `unknown`, access type `none`, interaction height `none`, surface map `front: [grasp, access_contents, open]`\n- low-confidence candidates: `none`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, group `unknown`, keywords [cocktail shaker], affordances [graspable, containable, openable], task targets [table, desk, nightstand, coffee_table, shelf, cabinet]\n- target `kitchen_base_cabinet_run_0`: category `cabinet`, group `storage`, keywords [cabinet, storage cabinet, cupboard], task targets [wall, clear_space]\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, room `kitchen`, placement `floor`, footprint `tiny`, height `very_low`, elevation `high`\nTarget cues:\n- target `kitchen_base_cabinet_run_0`: category `cabinet`, room `kitchen`, placement `floor`, footprint `large`, height `low`, elevation `floor_level`, subject_proximity `far`\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_salt_shaker_0`: category `cocktail shaker`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_print_3`: category `print_print`, placement `floor`, proximity `touching`, bbox_relation `very_close`, footprint `small`\n- `kitchen_floating_wall_shelf_0`: category `wall shelf`, placement `floor`, proximity `touching`, bbox_relation `very_close`, footprint `medium`\n- `kitchen_picture_frame_0`: category `picture frame`, placement `floor`, proximity `near`, bbox_relation `very_close`, footprint `small`\n- `kitchen_coffee_mug_3`: category `mug`, placement `floor`, proximity `near`, bbox_relation `very_close`, footprint `tiny`\n- `kitchen_potted_basil_herb_plant_0`: category `potted plant`, placement `floor`, proximity `near`, bbox_relation `nearby`, footprint `tiny`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `interaction_clearance` metric for `kitchen_pepper_shaker_0` with affordance `openable`.\nRelevant context: kitchen_base_cabinet_run_0.\nFirst infer the most likely operating side from door panels, handles, hinges, or the side that faces into the room. Then check only that side for a normal standing zone and immediate hand/door operating clearance. Do not assume a large hinged swing arc unless the render clearly shows hinged doors needing that arc. Do not require the entire front face to be empty; a nearby bench or seat off to one side is acceptable when the floor directly before the door/handle area remains usable. A narrow diagonal side gap or a squeeze-by path does not count as clear operating space.\nIf you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_pepper_shaker_0__to__kitchen_base_cabinet_run_0_local_operating_zone.png",
    "render_check_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_pepper_shaker_0__to__kitchen_base_cabinet_run_0_local_operating_zone_context.png",
    "render_check_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_pepper_shaker_0__to__kitchen_base_cabinet_run_0_local_depth_probe.png",
    "render_check_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/interaction_clearance__kitchen_pepper_shaker_0__to__kitchen_base_cabinet_run_0_local_plan_context.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view. 2. `local_context` is a highlighted context view; subject is red and targets are green.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: affordance_exposure
Subject object: kitchen_bistro_table_0
Target objects: kitchen_dining_chair_0, kitchen_dining_chair_1

Verified asset facts:
- subject `kitchen_bistro_table_0`: category `coffee table`, affordances [supportable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`
- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `top: [place_or_reach_objects]`
- low-confidence candidates: `none`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_bistro_table_0`: category `coffee table`, group `unknown`, keywords [coffee table], affordances [supportable], task targets [chair, refrigerator, room, graspable_object]
- target `kitchen_dining_chair_0`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], task targets [desk, table]
- target `kitchen_dining_chair_1`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], task targets [desk, table]

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_bistro_table_0`: category `coffee table`, room `kitchen`, placement `floor`, footprint `medium`, height `mid`, elevation `floor_level`
Target cues:
- target `kitchen_dining_chair_0`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`, subject_proximity `near`
- target `kitchen_dining_chair_1`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`, subject_proximity `near`
Nearby same-room context objects that may affect access or clearance:
- `kitchen_coffee_mug_0`: category `mug`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_coffee_mug_1`: category `mug`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`
- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_pendant_light_0`: category `pathway light`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `medium`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `affordance_exposure` metric for `kitchen_bistro_table_0` with affordance `supportable`.
Relevant context: kitchen_dining_chair_0, kitchen_dining_chair_1.
Check whether the usable top surface is visible enough to judge placement or reach. Do not require all table edges to be exposed.
If you choose `degraded` or `fail`, name at least one direct blocker or the directly constrained functional region. If no direct blocker is visible, prefer `pass` or `unknown`.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "affordance_exposure__kitchen_bistro_table_0__to__kitchen_dining_chair_0__kitchen_dining_chair_1",
  "metric": "affordance_exposure",
  "subject_id": "kitchen_bistro_table_0",
  "target_ids": [
    "kitchen_dining_chair_0",
    "kitchen_dining_chair_1"
  ],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view. 2. `local_context` is a highlighted context view; subject is red and targets are green.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: affordance_exposure\nSubject object: kitchen_bistro_table_0\nTarget objects: kitchen_dining_chair_0, kitchen_dining_chair_1\n\nVerified asset facts:\n- subject `kitchen_bistro_table_0`: category `coffee table`, affordances [supportable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`\n- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `top: [place_or_reach_objects]`\n- low-confidence candidates: `none`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_bistro_table_0`: category `coffee table`, group `unknown`, keywords [coffee table], affordances [supportable], task targets [chair, refrigerator, room, graspable_object]\n- target `kitchen_dining_chair_0`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], task targets [desk, table]\n- target `kitchen_dining_chair_1`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], task targets [desk, table]\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_bistro_table_0`: category `coffee table`, room `kitchen`, placement `floor`, footprint `medium`, height `mid`, elevation `floor_level`\nTarget cues:\n- target `kitchen_dining_chair_0`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`, subject_proximity `near`\n- target `kitchen_dining_chair_1`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`, subject_proximity `near`\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_coffee_mug_0`: category `mug`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_coffee_mug_1`: category `mug`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`\n- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_pendant_light_0`: category `pathway light`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `medium`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `affordance_exposure` metric for `kitchen_bistro_table_0` with affordance `supportable`.\nRelevant context: kitchen_dining_chair_0, kitchen_dining_chair_1.\nCheck whether the usable top surface is visible enough to judge placement or reach. Do not require all table edges to be exposed.\nIf you choose `degraded` or `fail`, name at least one direct blocker or the directly constrained functional region. If no direct blocker is visible, prefer `pass` or `unknown`.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/affordance_exposure__kitchen_bistro_table_0__to__kitchen_dining_chair_0__kitchen_dining_chair_1_local.png",
    "render_check_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/affordance_exposure__kitchen_bistro_table_0__to__kitchen_dining_chair_0__kitchen_dining_chair_1_local_context.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view. 2. `local_context` is a highlighted context view; subject is red and targets are green.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: spatial_accessibility
Subject object: kitchen_dining_chair_0
Target objects: none

Verified asset facts:
- subject `kitchen_dining_chair_0`: category `chair`, affordances [sittable, supportable, pushable, liftable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`
- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `front: [sit]; top: [place_or_reach_objects]`
- low-confidence candidates: `none`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_dining_chair_0`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], affordances [sittable, supportable, pushable, liftable], task targets [desk, table]
- targets: none

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_dining_chair_0`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`
Target cues:
- none
Nearby same-room context objects that may affect access or clearance:
- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_dining_chair_1`: category `chair`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `medium`
- `kitchen_eucalyptus_sprig_0_f0_1`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `small`
- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_eucalyptus_sprig_0_f0_2`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `spatial_accessibility` metric for `kitchen_dining_chair_0` with affordance `sittable`.
Relevant context: the surrounding scene.
Check whether a person can physically walk up to the seating edge and sit from the front or from one genuinely open side. Require one clearly usable approach corridor from open room space all the way to the seat edge itself, not merely to the furniture cluster or to an outer corner of the seat. A standard coffee table in front of a sofa or loveseat is normal unless the local renders show that it closes off the usable seating edge. Consider the combined effect of nearby furniture on the front and side approach zones only when that pinch is visually evident in the local renders. Close adjacency to a bed or wall is acceptable only if one normal sitting approach remains.
If you choose `degraded` or `fail`, name the blocker or the directly constrained approach or sit/stand zone. Consider the combined effect of nearby furniture only when the local renders clearly show that the front and side access are genuinely pinched or eliminated. A standard coffee table in front of a sofa or loveseat is normal unless the local evidence shows it closes off the usable seating edge. Do not infer exact centimeter clearance from auxiliary geometry or bounding-box numbers.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "spatial_accessibility__kitchen_dining_chair_0",
  "metric": "spatial_accessibility",
  "subject_id": "kitchen_dining_chair_0",
  "target_ids": [],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view. 2. `local_context` is a highlighted context view; subject is red and targets are green.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: spatial_accessibility\nSubject object: kitchen_dining_chair_0\nTarget objects: none\n\nVerified asset facts:\n- subject `kitchen_dining_chair_0`: category `chair`, affordances [sittable, supportable, pushable, liftable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`\n- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `front: [sit]; top: [place_or_reach_objects]`\n- low-confidence candidates: `none`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_dining_chair_0`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], affordances [sittable, supportable, pushable, liftable], task targets [desk, table]\n- targets: none\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_dining_chair_0`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`\nTarget cues:\n- none\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_dining_chair_1`: category `chair`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `medium`\n- `kitchen_eucalyptus_sprig_0_f0_1`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `small`\n- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_eucalyptus_sprig_0_f0_2`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `spatial_accessibility` metric for `kitchen_dining_chair_0` with affordance `sittable`.\nRelevant context: the surrounding scene.\nCheck whether a person can physically walk up to the seating edge and sit from the front or from one genuinely open side. Require one clearly usable approach corridor from open room space all the way to the seat edge itself, not merely to the furniture cluster or to an outer corner of the seat. A standard coffee table in front of a sofa or loveseat is normal unless the local renders show that it closes off the usable seating edge. Consider the combined effect of nearby furniture on the front and side approach zones only when that pinch is visually evident in the local renders. Close adjacency to a bed or wall is acceptable only if one normal sitting approach remains.\nIf you choose `degraded` or `fail`, name the blocker or the directly constrained approach or sit/stand zone. Consider the combined effect of nearby furniture only when the local renders clearly show that the front and side access are genuinely pinched or eliminated. A standard coffee table in front of a sofa or loveseat is normal unless the local evidence shows it closes off the usable seating edge. Do not infer exact centimeter clearance from auxiliary geometry or bounding-box numbers.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_dining_chair_0_local.png",
    "render_check_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_dining_chair_0_local_context.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: spatial_accessibility
Subject object: kitchen_dining_chair_1
Target objects: none

Verified asset facts:
- subject `kitchen_dining_chair_1`: category `chair`, affordances [sittable, supportable, pushable, liftable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`
- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `front: [sit]; top: [place_or_reach_objects]`
- low-confidence candidates: `none`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_dining_chair_1`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], affordances [sittable, supportable, pushable, liftable], task targets [desk, table]
- targets: none

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_dining_chair_1`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`
Target cues:
- none
Nearby same-room context objects that may affect access or clearance:
- `kitchen_dining_chair_0`: category `chair`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `medium`
- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `small`
- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_eucalyptus_sprig_0_f0_0`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_eucalyptus_sprig_0_f0_2`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `spatial_accessibility` metric for `kitchen_dining_chair_1` with affordance `sittable`.
Relevant context: the surrounding scene.
Check whether a person can physically walk up to the seating edge and sit from the front or from one genuinely open side. Require one clearly usable approach corridor from open room space all the way to the seat edge itself, not merely to the furniture cluster or to an outer corner of the seat. A standard coffee table in front of a sofa or loveseat is normal unless the local renders show that it closes off the usable seating edge. Consider the combined effect of nearby furniture on the front and side approach zones only when that pinch is visually evident in the local renders. Close adjacency to a bed or wall is acceptable only if one normal sitting approach remains.
If you choose `degraded` or `fail`, name the blocker or the directly constrained approach or sit/stand zone. Consider the combined effect of nearby furniture only when the local renders clearly show that the front and side access are genuinely pinched or eliminated. A standard coffee table in front of a sofa or loveseat is normal unless the local evidence shows it closes off the usable seating edge. Do not infer exact centimeter clearance from auxiliary geometry or bounding-box numbers.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "spatial_accessibility__kitchen_dining_chair_1",
  "metric": "spatial_accessibility",
  "subject_id": "kitchen_dining_chair_1",
  "target_ids": [],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: spatial_accessibility\nSubject object: kitchen_dining_chair_1\nTarget objects: none\n\nVerified asset facts:\n- subject `kitchen_dining_chair_1`: category `chair`, affordances [sittable, supportable, pushable, liftable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`\n- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `front: [sit]; top: [place_or_reach_objects]`\n- low-confidence candidates: `none`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_dining_chair_1`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], affordances [sittable, supportable, pushable, liftable], task targets [desk, table]\n- targets: none\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_dining_chair_1`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`\nTarget cues:\n- none\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_dining_chair_0`: category `chair`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `medium`\n- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `small`\n- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_eucalyptus_sprig_0_f0_0`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_eucalyptus_sprig_0_f0_2`: category `eucalyptus_sprig_f0`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `spatial_accessibility` metric for `kitchen_dining_chair_1` with affordance `sittable`.\nRelevant context: the surrounding scene.\nCheck whether a person can physically walk up to the seating edge and sit from the front or from one genuinely open side. Require one clearly usable approach corridor from open room space all the way to the seat edge itself, not merely to the furniture cluster or to an outer corner of the seat. A standard coffee table in front of a sofa or loveseat is normal unless the local renders show that it closes off the usable seating edge. Consider the combined effect of nearby furniture on the front and side approach zones only when that pinch is visually evident in the local renders. Close adjacency to a bed or wall is acceptable only if one normal sitting approach remains.\nIf you choose `degraded` or `fail`, name the blocker or the directly constrained approach or sit/stand zone. Consider the combined effect of nearby furniture only when the local renders clearly show that the front and side access are genuinely pinched or eliminated. A standard coffee table in front of a sofa or loveseat is normal unless the local evidence shows it closes off the usable seating edge. Do not infer exact centimeter clearance from auxiliary geometry or bounding-box numbers.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_dining_chair_1_local.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: spatial_accessibility
Subject object: kitchen_fridge_0
Target objects: none

Verified asset facts:
- subject `kitchen_fridge_0`: no offline asset YAML loaded; heuristic category `refrigerator`, affordances [openable, containable], usable face `front`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_fridge_0`: category `refrigerator`, group `appliance_storage`, keywords [refrigerator, fridge, cooling cabinet], affordances [openable, containable], task targets [wall, clear_space]
- targets: none

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_fridge_0`: category `refrigerator`, room `kitchen`, placement `floor`, footprint `large`, height `mid`, elevation `floor_level`
Target cues:
- none
Nearby same-room context objects that may affect access or clearance:
- `kitchen_semi_flush_ceiling_light_0`: category `pathway light`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`
- `kitchen_kitchen_prep_cart_0`: category `serving cart`, placement `floor`, proximity `mid`, bbox_relation `overlapping_or_touching`, footprint `medium`
- `kitchen_canister_jar_0`: category `jar`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`
- `kitchen_kitchen_towel_0`: category `towel`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`
- `kitchen_teaspoon_0`: category `teaspoon_teaspoon`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`
- `kitchen_chefs_knife_0_s1_1`: category `chefs_knife_s1`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `small`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `spatial_accessibility` metric for `kitchen_fridge_0` with affordance `openable`.
Relevant context: the surrounding scene.
Check whether at least one practical front-facing or slightly offset standing position can reach the operating side so the object can be opened. A nearby bed corner or bench should not cause degraded by itself if a normal reach from the front region is still possible. Use degraded only when that front-region approach is directly pinched, occupied, or effectively unusable across the usable operating side. Do not count a cramped diagonal squeeze, shoulder-turned sidestep, or tiny side sliver as a valid approach path.
If you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "spatial_accessibility__kitchen_fridge_0",
  "metric": "spatial_accessibility",
  "subject_id": "kitchen_fridge_0",
  "target_ids": [],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: spatial_accessibility\nSubject object: kitchen_fridge_0\nTarget objects: none\n\nVerified asset facts:\n- subject `kitchen_fridge_0`: no offline asset YAML loaded; heuristic category `refrigerator`, affordances [openable, containable], usable face `front`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_fridge_0`: category `refrigerator`, group `appliance_storage`, keywords [refrigerator, fridge, cooling cabinet], affordances [openable, containable], task targets [wall, clear_space]\n- targets: none\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_fridge_0`: category `refrigerator`, room `kitchen`, placement `floor`, footprint `large`, height `mid`, elevation `floor_level`\nTarget cues:\n- none\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_semi_flush_ceiling_light_0`: category `pathway light`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`\n- `kitchen_kitchen_prep_cart_0`: category `serving cart`, placement `floor`, proximity `mid`, bbox_relation `overlapping_or_touching`, footprint `medium`\n- `kitchen_canister_jar_0`: category `jar`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`\n- `kitchen_kitchen_towel_0`: category `towel`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`\n- `kitchen_teaspoon_0`: category `teaspoon_teaspoon`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `tiny`\n- `kitchen_chefs_knife_0_s1_1`: category `chefs_knife_s1`, placement `floor`, proximity `mid`, bbox_relation `nearby`, footprint `small`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `spatial_accessibility` metric for `kitchen_fridge_0` with affordance `openable`.\nRelevant context: the surrounding scene.\nCheck whether at least one practical front-facing or slightly offset standing position can reach the operating side so the object can be opened. A nearby bed corner or bench should not cause degraded by itself if a normal reach from the front region is still possible. Use degraded only when that front-region approach is directly pinched, occupied, or effectively unusable across the usable operating side. Do not count a cramped diagonal squeeze, shoulder-turned sidestep, or tiny side sliver as a valid approach path.\nIf you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_fridge_0_local_operating_zone.png",
    "render_check_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_fridge_0_local_operating_zone_context.png",
    "render_check_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_fridge_0_local_depth_probe.png",
    "render_check_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_fridge_0_local_plan_context.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view. 2. `local_context` is a highlighted context view; subject is red and targets are green.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: spatial_accessibility
Subject object: kitchen_bistro_table_0
Target objects: kitchen_dining_chair_0, kitchen_dining_chair_1

Verified asset facts:
- subject `kitchen_bistro_table_0`: category `coffee table`, affordances [supportable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`
- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `top: [place_or_reach_objects]`
- low-confidence candidates: `none`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_bistro_table_0`: category `coffee table`, group `unknown`, keywords [coffee table], affordances [supportable], task targets [chair, refrigerator, room, graspable_object]
- target `kitchen_dining_chair_0`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], task targets [desk, table]
- target `kitchen_dining_chair_1`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], task targets [desk, table]

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_bistro_table_0`: category `coffee table`, room `kitchen`, placement `floor`, footprint `medium`, height `mid`, elevation `floor_level`
Target cues:
- target `kitchen_dining_chair_0`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`, subject_proximity `near`
- target `kitchen_dining_chair_1`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`, subject_proximity `near`
Nearby same-room context objects that may affect access or clearance:
- `kitchen_coffee_mug_0`: category `mug`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_coffee_mug_1`: category `mug`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`
- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_pendant_light_0`: category `pathway light`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `medium`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `spatial_accessibility` metric for `kitchen_bistro_table_0` with affordance `supportable`.
Relevant context: kitchen_dining_chair_0, kitchen_dining_chair_1.
Check whether a person can approach the room-facing side and reach the usable top surface. Do not require circulation on all sides.
If you choose `degraded` or `fail`, name at least one direct blocker or the directly constrained functional region. If no direct blocker is visible, prefer `pass` or `unknown`.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "spatial_accessibility__kitchen_bistro_table_0__to__kitchen_dining_chair_0__kitchen_dining_chair_1",
  "metric": "spatial_accessibility",
  "subject_id": "kitchen_bistro_table_0",
  "target_ids": [
    "kitchen_dining_chair_0",
    "kitchen_dining_chair_1"
  ],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local` is the main clean RGB local inspection view. 2. `local_context` is a highlighted context view; subject is red and targets are green.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: spatial_accessibility\nSubject object: kitchen_bistro_table_0\nTarget objects: kitchen_dining_chair_0, kitchen_dining_chair_1\n\nVerified asset facts:\n- subject `kitchen_bistro_table_0`: category `coffee table`, affordances [supportable], usable face `top`, benchmark relevance `functional`, confidence `0.67`, source `asset_annotation`\n- interaction facts: placement class `floor_furniture`, access type `none`, interaction height `none`, surface map `top: [place_or_reach_objects]`\n- low-confidence candidates: `none`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_bistro_table_0`: category `coffee table`, group `unknown`, keywords [coffee table], affordances [supportable], task targets [chair, refrigerator, room, graspable_object]\n- target `kitchen_dining_chair_0`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], task targets [desk, table]\n- target `kitchen_dining_chair_1`: category `chair`, group `seating`, keywords [chair, side chair, dining chair, seating], task targets [desk, table]\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_bistro_table_0`: category `coffee table`, room `kitchen`, placement `floor`, footprint `medium`, height `mid`, elevation `floor_level`\nTarget cues:\n- target `kitchen_dining_chair_0`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`, subject_proximity `near`\n- target `kitchen_dining_chair_1`: category `chair`, room `kitchen`, placement `floor`, footprint `medium`, height `low`, elevation `floor_level`, subject_proximity `near`\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_coffee_mug_0`: category `mug`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_coffee_mug_1`: category `mug`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_croissant_pastry_0_s0_1`: category `croissant_pastry_s0`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_dessert_plate_0_s0_0`: category `plate`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`\n- `kitchen_bud_vase_0_f0_c`: category `vase`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_pendant_light_0`: category `pathway light`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `medium`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `spatial_accessibility` metric for `kitchen_bistro_table_0` with affordance `supportable`.\nRelevant context: kitchen_dining_chair_0, kitchen_dining_chair_1.\nCheck whether a person can approach the room-facing side and reach the usable top surface. Do not require circulation on all sides.\nIf you choose `degraded` or `fail`, name at least one direct blocker or the directly constrained functional region. If no direct blocker is visible, prefer `pass` or `unknown`.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_bistro_table_0__to__kitchen_dining_chair_0__kitchen_dining_chair_1_local.png",
    "render_check_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_bistro_table_0__to__kitchen_dining_chair_0__kitchen_dining_chair_1_local_context.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: spatial_accessibility
Subject object: kitchen_base_cabinet_run_0
Target objects: none

Verified asset facts:
- subject `kitchen_base_cabinet_run_0`: no offline asset YAML loaded; heuristic category `cabinet`, affordances [openable, containable, supportable], usable face `top`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_base_cabinet_run_0`: category `cabinet`, group `storage`, keywords [cabinet, storage cabinet, cupboard], affordances [openable, containable, supportable], task targets [wall, clear_space]
- targets: none

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_base_cabinet_run_0`: category `cabinet`, room `kitchen`, placement `floor`, footprint `large`, height `low`, elevation `floor_level`
Target cues:
- none
Nearby same-room context objects that may affect access or clearance:
- `kitchen_hand_soap_dispenser_0`: category `soap dispenser`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_kitchen_prep_cart_0`: category `serving cart`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `medium`
- `kitchen_dish_drying_mat_0`: category `place mat`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`
- `kitchen_dish_soap_pump_bottle_0`: category `bottle`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_kitchen_sponge_scrubber_0`: category `sponge_scrubber_sponge`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_electric_kettle_1`: category `kettle`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `spatial_accessibility` metric for `kitchen_base_cabinet_run_0` with affordance `openable`.
Relevant context: the surrounding scene.
Check whether at least one practical front-facing or slightly offset standing position can reach the operating side so the object can be opened. A nearby bed corner or bench should not cause degraded by itself if a normal reach from the front region is still possible. Use degraded only when that front-region approach is directly pinched, occupied, or effectively unusable across the usable operating side. Do not count a cramped diagonal squeeze, shoulder-turned sidestep, or tiny side sliver as a valid approach path.
If you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "spatial_accessibility__kitchen_base_cabinet_run_0",
  "metric": "spatial_accessibility",
  "subject_id": "kitchen_base_cabinet_run_0",
  "target_ids": [],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: spatial_accessibility\nSubject object: kitchen_base_cabinet_run_0\nTarget objects: none\n\nVerified asset facts:\n- subject `kitchen_base_cabinet_run_0`: no offline asset YAML loaded; heuristic category `cabinet`, affordances [openable, containable, supportable], usable face `top`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_base_cabinet_run_0`: category `cabinet`, group `storage`, keywords [cabinet, storage cabinet, cupboard], affordances [openable, containable, supportable], task targets [wall, clear_space]\n- targets: none\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_base_cabinet_run_0`: category `cabinet`, room `kitchen`, placement `floor`, footprint `large`, height `low`, elevation `floor_level`\nTarget cues:\n- none\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_hand_soap_dispenser_0`: category `soap dispenser`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_kitchen_prep_cart_0`: category `serving cart`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `medium`\n- `kitchen_dish_drying_mat_0`: category `place mat`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`\n- `kitchen_dish_soap_pump_bottle_0`: category `bottle`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_kitchen_sponge_scrubber_0`: category `sponge_scrubber_sponge`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_electric_kettle_1`: category `kettle`, placement `floor`, proximity `near`, bbox_relation `overlapping_or_touching`, footprint `small`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `spatial_accessibility` metric for `kitchen_base_cabinet_run_0` with affordance `openable`.\nRelevant context: the surrounding scene.\nCheck whether at least one practical front-facing or slightly offset standing position can reach the operating side so the object can be opened. A nearby bed corner or bench should not cause degraded by itself if a normal reach from the front region is still possible. Use degraded only when that front-region approach is directly pinched, occupied, or effectively unusable across the usable operating side. Do not count a cramped diagonal squeeze, shoulder-turned sidestep, or tiny side sliver as a valid approach path.\nIf you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_base_cabinet_run_0_local_operating_zone.png",
    "render_check_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_base_cabinet_run_0_local_operating_zone_context.png",
    "render_check_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_base_cabinet_run_0_local_depth_probe.png",
    "render_check_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_base_cabinet_run_0_local_plan_context.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}

You are an affordance-aware functional critic for 3D indoor scenes.

Core principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.

Workflow:
1. Read the task instruction and target metric.
2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.
3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.
4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.
5. Evaluate ONLY the requested metric and return a JSON object matching the schema.

Scene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.
Task instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Metric: spatial_accessibility
Subject object: kitchen_pepper_shaker_0
Target objects: kitchen_base_cabinet_run_0

Verified asset facts:
- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, affordances [graspable, containable, openable], usable face `front`, benchmark relevance `functional`, confidence `0.58`, source `asset_annotation`
- interaction facts: placement class `unknown`, access type `none`, interaction height `none`, surface map `front: [grasp, access_contents, open]`
- low-confidence candidates: `none`
Use these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.

Scene evidence object hints:
- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, group `unknown`, keywords [cocktail shaker], affordances [graspable, containable, openable], task targets [table, desk, nightstand, coffee_table, shelf, cabinet]
- target `kitchen_base_cabinet_run_0`: category `cabinet`, group `storage`, keywords [cabinet, storage cabinet, cupboard], task targets [wall, clear_space]

Auxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):
- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, room `kitchen`, placement `floor`, footprint `tiny`, height `very_low`, elevation `high`
Target cues:
- target `kitchen_base_cabinet_run_0`: category `cabinet`, room `kitchen`, placement `floor`, footprint `large`, height `low`, elevation `floor_level`, subject_proximity `far`
Nearby same-room context objects that may affect access or clearance:
- `kitchen_salt_shaker_0`: category `cocktail shaker`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`
- `kitchen_print_3`: category `print_print`, placement `floor`, proximity `touching`, bbox_relation `very_close`, footprint `small`
- `kitchen_floating_wall_shelf_0`: category `wall shelf`, placement `floor`, proximity `touching`, bbox_relation `very_close`, footprint `medium`
- `kitchen_picture_frame_0`: category `picture frame`, placement `floor`, proximity `near`, bbox_relation `very_close`, footprint `small`
- `kitchen_coffee_mug_3`: category `mug`, placement `floor`, proximity `near`, bbox_relation `very_close`, footprint `tiny`
- `kitchen_potted_basil_herb_plant_0`: category `potted plant`, placement `floor`, proximity `near`, bbox_relation `nearby`, footprint `tiny`
Do not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.

Metric-specific rule:
Task: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.
Evaluate only the `spatial_accessibility` metric for `kitchen_pepper_shaker_0` with affordance `openable`.
Relevant context: kitchen_base_cabinet_run_0.
Check whether at least one practical front-facing or slightly offset standing position can reach the operating side so the object can be opened. A nearby bed corner or bench should not cause degraded by itself if a normal reach from the front region is still possible. Use degraded only when that front-region approach is directly pinched, occupied, or effectively unusable across the usable operating side. Do not count a cramped diagonal squeeze, shoulder-turned sidestep, or tiny side sliver as a valid approach path.
If you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.

Image order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.
Evidence guardrails:
- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.
- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.
- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.
- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.
- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.
Output JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.
Allowed labels: pass, degraded, fail, unknown.
2D perspective overlap alone is not enough evidence of a blocker.
Normal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.
If you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.
If evidence is insufficient, use unknown instead of guessing.

{
  "check_id": "spatial_accessibility__kitchen_pepper_shaker_0__to__kitchen_base_cabinet_run_0",
  "metric": "spatial_accessibility",
  "subject_id": "kitchen_pepper_shaker_0",
  "target_ids": [
    "kitchen_base_cabinet_run_0"
  ],
  "prompt_text": "You are an affordance-aware functional critic for 3D indoor scenes.\n\nCore principle: judge whether the task CAN be done in the current scene, without assuming it has already been pre-solved.\n\nWorkflow:\n1. Read the task instruction and target metric.\n2. Inspect the local check renders first and treat them as the primary evidence for the requested interaction.\n3. Local render order and intended use: 1. `local_operating_zone` is the clean RGB view for the most likely door-facing standing and opening zone. 2. `local_operating_zone_context` is a slightly wider highlighted context view of the same operating side. 3. `local_depth_probe` is an oblique RGB view meant to clarify depth and whether a blocker sits in front of the operating zone. 4. `local_plan_context` is a high local layout view used to check whether a nearby object actually spans or occupies the operating/approach zone rather than merely overlapping in perspective.\n4. Then inspect the 4 attached overview renders for room-scale layout and approach context. These overviews are captured from a slightly elevated height and are not meant to mimic adult eye level.\n5. Evaluate ONLY the requested metric and return a JSON object matching the schema.\n\nScene summary: scenesmith scene with 51 objects. Room type: kitchen. Planned checks: 12. Room shell: 3.80m x 3.40m x 2.70m.\nTask instruction: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nMetric: spatial_accessibility\nSubject object: kitchen_pepper_shaker_0\nTarget objects: kitchen_base_cabinet_run_0\n\nVerified asset facts:\n- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, affordances [graspable, containable, openable], usable face `front`, benchmark relevance `functional`, confidence `0.58`, source `asset_annotation`\n- interaction facts: placement class `unknown`, access type `none`, interaction height `none`, surface map `front: [grasp, access_contents, open]`\n- low-confidence candidates: `none`\nUse these as converter-side asset facts. Do not reclassify the subject; judge only whether the listed affordance remains usable in the current scene. If the image clearly contradicts these facts, return unknown and explain the conflict.\n\nScene evidence object hints:\n- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, group `unknown`, keywords [cocktail shaker], affordances [graspable, containable, openable], task targets [table, desk, nightstand, coffee_table, shelf, cabinet]\n- target `kitchen_base_cabinet_run_0`: category `cabinet`, group `storage`, keywords [cabinet, storage cabinet, cupboard], task targets [wall, clear_space]\n\nAuxiliary scene cues (coarse geometry only; use these cues just to identify rough scale and nearby context):\n- subject `kitchen_pepper_shaker_0`: category `cocktail shaker`, room `kitchen`, placement `floor`, footprint `tiny`, height `very_low`, elevation `high`\nTarget cues:\n- target `kitchen_base_cabinet_run_0`: category `cabinet`, room `kitchen`, placement `floor`, footprint `large`, height `low`, elevation `floor_level`, subject_proximity `far`\nNearby same-room context objects that may affect access or clearance:\n- `kitchen_salt_shaker_0`: category `cocktail shaker`, placement `floor`, proximity `touching`, bbox_relation `overlapping_or_touching`, footprint `tiny`\n- `kitchen_print_3`: category `print_print`, placement `floor`, proximity `touching`, bbox_relation `very_close`, footprint `small`\n- `kitchen_floating_wall_shelf_0`: category `wall shelf`, placement `floor`, proximity `touching`, bbox_relation `very_close`, footprint `medium`\n- `kitchen_picture_frame_0`: category `picture frame`, placement `floor`, proximity `near`, bbox_relation `very_close`, footprint `small`\n- `kitchen_coffee_mug_3`: category `mug`, placement `floor`, proximity `near`, bbox_relation `very_close`, footprint `tiny`\n- `kitchen_potted_basil_herb_plant_0`: category `potted plant`, placement `floor`, proximity `near`, bbox_relation `nearby`, footprint `tiny`\nDo not estimate exact centimeter clearance from these coarse cues. They are only for object identity, rough scale, and coarse proximity; the local renders remain the deciding evidence.\n\nMetric-specific rule:\nTask: A kitchen featuring a bistro table with two chairs in a corner and a refrigerator positioned nearby.\nEvaluate only the `spatial_accessibility` metric for `kitchen_pepper_shaker_0` with affordance `openable`.\nRelevant context: kitchen_base_cabinet_run_0.\nCheck whether at least one practical front-facing or slightly offset standing position can reach the operating side so the object can be opened. A nearby bed corner or bench should not cause degraded by itself if a normal reach from the front region is still possible. Use degraded only when that front-region approach is directly pinched, occupied, or effectively unusable across the usable operating side. Do not count a cramped diagonal squeeze, shoulder-turned sidestep, or tiny side sliver as a valid approach path.\nIf you choose `degraded` or `fail`, name the blocker or the blocked standing/opening zone directly in front of the operating side. If at least one practical front-facing or slightly offset standing position remains and can still reach the operating side normally, prefer `pass`. Do not fail an openable object merely because a nearby seat or small furniture item is offset near one part of the front; require evidence that the usable operating side is actually occupied or pinched. If the operating side cannot be identified from the evidence, return `unknown` instead of guessing.\n\nImage order is fixed: local check renders always come first, followed by the overview renders. If a local render is highlighted, the subject is red and the targets are green.\nEvidence guardrails:\n- Treat the local renders as the deciding evidence for direct blockage, operating-side clearance, and immediate approach space.\n- Use overview renders only to understand room-scale layout or to break ties when the local evidence is incomplete.\n- To output `degraded` or `fail`, the local evidence should explicitly show a directly constrained operating/approach zone, not just a nearby object or a single perspective overlap.\n- When a `local_plan_context` view is present, use it to verify whether the suspected blocker really spans the usable operating/approach zone; if it shows a normal nearby standing/reach position remains, prefer `pass`.\n- If the overview render suggests a blockage but the local evidence does not clearly show that blockage, prefer `pass` or `unknown`, not `degraded` or `fail`.\nOutput JSON with keys: check_id, metric, label, asset_fact_used, asset_fact_confidence, evidence_conflict, reason, blocking_objects, confidence.\nAllowed labels: pass, degraded, fail, unknown.\n2D perspective overlap alone is not enough evidence of a blocker.\nNormal adjacency priors should not override direct evidence in the local renders. Nightstands beside beds, benches touching bed ends, corner wardrobes, and wall decor are acceptable only when they do not directly intrude into the required functional region.\nIf you choose degraded or fail, the reason must name a direct blocker or directly constrained functional region.\nIf evidence is insufficient, use unknown instead of guessing.",
  "files": {
    "render_check_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_pepper_shaker_0__to__kitchen_base_cabinet_run_0_local_operating_zone.png",
    "render_check_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_pepper_shaker_0__to__kitchen_base_cabinet_run_0_local_operating_zone_context.png",
    "render_check_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_pepper_shaker_0__to__kitchen_base_cabinet_run_0_local_depth_probe.png",
    "render_check_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/checks/spatial_accessibility__kitchen_pepper_shaker_0__to__kitchen_base_cabinet_run_0_local_plan_context.png",
    "render_overview_image": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_00_rgb.png",
    "render_overview_image_01": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_01_rgb.png",
    "render_overview_image_02": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_02_rgb.png",
    "render_overview_image_03": "/data/250010098/SceneBenchmark/res/vlm_runs/scenesmith_ud4/scenesmith-example-data_scene_064/render_bundle/views/view_03_rgb.png"
  }
}