task_three_d__object_cluster__counterfactual_count

Summary

• Domain: three_d
• Scene id: object_cluster
• Package: src/trace_tasks/tasks/three_d/object_cluster/
• Supported query_id: single
• Answer type: integer
• Annotation type: unordered bbox_set
• Annotation schema: bbox_set

Program Contract

Program: count(filter(object_cluster_objects, predicate = target_predicate)) + sum(edit_delta_i(target_predicate, ordered_edit_sequence)); scene=object_cluster; scope=counterfactual_count

Candidate set: the visible 3D objects, surfaces, room/street/warehouse structures, spatial anchors, markers, and labeled options inside the counterfactual_count objective scope. Operands: visible scene state and prompt-bound operands named by filter, object_cluster_objects, predicate, target_predicate, sum, edit_delta_i, ordered_edit_sequence, object_cluster, counterfactual_count. Operation: evaluate count 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 integer schema; generation binds a unique final answer. Annotation witnesses: annotation uses the bbox_set schema; the prompt/annotation contract defines the minimal visual witnesses. Query ids: single.

Reasoning Operations

Families: filtering, counting, aggregation, state_update, formula_evaluation

Contract

The image shows many small synthetic perspective 3D objects arranged on a plain surface. This scene is a bare clustered-counting surface: it does not use option labels, named reference objects, relation prompts, or grid-based spatial cues.

The prompt asks for the final count of a queried property after applying an ordered sequence of textual add/remove edits to the visible starting cluster. The target predicate is one of:

• object type only, such as cups
• semantic color only, such as blue [#1F77B4] objects
• object type plus semantic color, such as blue [#1F77B4] cups

Each instance uses 2-3 ordered edits. At least one edit affects the queried property and at least one edit is a distractor whose predicate is disjoint from the queried property. Three-step instances may include a second target-affecting edit. Remove edits are generated only when enough starting visible objects matching that edit predicate are available, and generation keeps the final answer positive. Semantic color targets use the repo-wide canonical named color palette and prompt-facing hex labels.

The image includes non-target distractors. For color+object targets, generation includes structured partial-match distractors when possible, such as same-type/wrong-color and same-color/wrong-type objects. Wrong-type distractors avoid target-confusable object families such as card/envelope/book, sphere/button, cup/bowl/tray, lantern/candle, and pencil/ruler. The answer is computed from metadata as:

initial_visible_target_count + sum(target_affecting_add_amounts) - sum(target_affecting_remove_amounts)

Pixels are render output, not verifier source of truth.

Annotation Contract

Annotation is a bbox_set containing one [x0, y0, x1, y1] pixel box around each starting visible object matching the queried property before the edits. The final answer can differ from the annotation cardinality because the prompt asks for the count after the hypothetical edit sequence.

The annotation set is unordered because all visible witnesses have the same semantic role.

Prompt And Trace

The prompt bundle is three_d_object_cluster_v1 under src/trace_tasks/resources/prompts/three_d/object_cluster/. The trace records camera pose, projection frame, object world coordinates, sampled dimensions, prompt-facing object names, semantic colors, target predicate kind, initial target count, ordered counterfactual edit steps, target-affecting deltas, final target count, target object ids, projected object boxes, and the solver count predicate.

Determinism

Generation is deterministic from instance_seed, explicit params, config defaults, prompt bundle, and code versions. The starting visual annotation and final counterfactual answer are recorded in the same finalized 3D scene trace.