Skip to content

task_icons__named_field__multi_attribute_xor_count

Identity

  • domain: icons
  • scene_id: named_field
  • task: multi_attribute_xor
  • module: src/trace_tasks/tasks/icons/named_field/multi_attribute_xor_count.py
  • prompt bundle: src/trace_tasks/resources/prompts/icons/named_field/icons_named_field_v1.json

Program Contract

Program: count.multi_attribute_xor(scene=named_field, scope=all_icons, predicates=exactly_one_of_shape_or_secondary_attribute, secondary_attribute=color, output=count)

Candidate set: the visible icon instances, icon attributes, fields, grids, paths, panels, reference items, and labeled option cards inside the all_icons objective scope. Operands: visible scene state and prompt-bound operands named by named_field, all_icons, predicates, exactly_one_of_shape_or_secondary_attribute, secondary_attribute, color. Operation: evaluate count.multi_attribute_xor over the candidate set using the visible icon attributes, positions, relationships, transforms, counts, comparisons, 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, logical_composition

Scene And Query

The task renders one panel containing procedurally generated named shape icons. Each icon has a semantic procedural shape_id and a semantic color_name sampled from the shared Trace named-color palette. Icons may also render with non-semantic fill patterns (solid, striped, or dotted) as visual variation, but fill style is not queried by this task.

The prompt names one shape and one color, then asks for the count satisfying one Boolean predicate.

Supported public query ids: - single

The fixed internal predicate key is exactly_one_shape_or_color_count; trace metadata records it as query_spec.internal_query_id and query_spec.params.internal_query_id. It selects the exactly-one-of-shape-or-color predicate.

The target shape support is the full procedural named-icon vocabulary in src/trace_tasks/tasks/icons/shared/procedural_named_icons.py. Color prompts use the shared named-color label with hex notation, such as red [#E63232], from src/trace_tasks/tasks/shared/named_colors.py.

The scene uses only non-stack named-icon arrangements: jittered/ordered grids, shelf rows, collision-free scatter, and clusters by shape. Stack arrangements are intentionally excluded because Boolean shape/color membership should be answered by attribute filtering rather than row/column stack arithmetic.

Answer Contract

  • answer_gt.type = integer
  • value is the count of icons satisfying the exactly-one-of-shape-or-color predicate
  • default answer support is 1..5, with lower counts sampled more often to keep the generated answer distribution stable

Annotation Contract

  • annotation_gt.type = bbox_set
  • one [x0, y0, x1, y1] pixel bounding box for each counted icon
  • annotation boxes are sorted by the witnesses in reading order

Trace Contract

  • scene_ir.entities contains one entity for each rendered procedural icon.
  • Each entity records shape_id, prompt-facing shape_name, color_name, fill_style, RGB tint, bbox, size, rotation, placement group, row/column layout coordinates, and icon-noise metadata.
  • query_spec.params.query_id is single.
  • query_spec.params.internal_query_id records the fixed Boolean predicate key.
  • query_spec.params.target_attribute_axis records whether the secondary attribute is color or fill_style.
  • query_spec.params.partition_counts records the four target-relative partitions: both, shape_only, attribute_only, and neither.
  • execution_trace.target_answer equals the integer answer.
  • render_map.counted_instance_ids, witness_symbolic.counted_instance_ids, and projected_annotation.bbox_set are derived from the same rendered instances.

Prompt Contract

  • scene_key = single_scene_counting
  • task_key = counting_query
  • prompts ask for the count of one named shape plus color-or-fill-style Boolean condition
  • Both output modes include contract-valid JSON examples