task_three_d__object_scene__reference_nearest_label

Summary

• Domain: three_d
• Scene id: object_scene
• Package: src/trace_tasks/tasks/three_d/object_scene/
• Supported query_id: closest_to_reference, farthest_from_reference
• Answer type: option_letter
• Annotation type: bbox
• Annotation schema: bbox

Program Contract

Program: select(label(candidate_objects, extremum(projected_screen_center_distance_to_reference))); scene=object_scene; scope=reference_nearest_label

Candidate set: the visible 3D objects, surfaces, room/street/warehouse structures, spatial anchors, markers, and labeled options inside the reference_nearest_label objective scope. Operands: visible scene state and prompt-bound operands named by label, candidate_objects, extremum, projected_screen_center_distance_to_reference, object_scene, reference_nearest_label plus the active query_id branch. Operation: evaluate select over the candidate set using the finalized 3D scene state, camera projection, object identities, spatial relations, counts, distances, or option-selection constraints encoded in the program expression; generation enforces a unique final answer. Output binding: answer uses the option_letter schema; intentionally based on projected visual distance because this is what. Annotation witnesses: annotation uses the bbox schema; the prompt/annotation contract defines the minimal visual witnesses. Query ids: closest_to_reference, farthest_from_reference.

Reasoning Operations

Families: ranking, spatial_relations

Contract

The image uses the object_scene renderer: a perspective 3D floor, table, or platform scene with projected objects, markers, references, or paired views depending on the task. The public task id defines the stable objective contract; query ids are used only for genuine semantic operations within that contract. Render style, camera, canvas preset, object placement, labels, colors, and prompt wording variants are generation metadata, not public task axes.

The verifier computes the answer from finalized scene metadata and projection records, not from pixels. The prompt bundle is three_d_object_scene_v1 under src/trace_tasks/resources/prompts/three_d/object_scene/.

Each scene contains exactly one large named reference prop and six small named candidate objects. Candidate objects are placed in front of or beside the reference from the camera view, with render-time rejection if a candidate is hidden behind the reference prop. Query ids choose closest vs farthest by projected screen-space center distance to the reference.

The trace also records 3D surface-gap ordering as diagnostic metadata, but the answer is intentionally based on projected visual distance because this is what the rendered question asks the model to judge. The selected answer must have a clear screen-space margin. The task excludes heart, sword, and remote_control candidates because their silhouettes made the visual relation ambiguous during generation checks.

Annotation Contract

Annotation is a scalar bbox around the selected visible object. The selected object is the only visual witness; option text is not annotation.

Prompt And Trace

The trace records selected prompt keys, camera/projection data, object or marker records, rendered pixel witnesses, answer-support metadata, and the solver fields needed to recompute the answer and annotation from the same finalized scene.

Determinism

Generation is deterministic from instance_seed, explicit params, config defaults, prompt bundle, and code versions. Answers and annotation come from the same finalized 3D scene trace.