`task_icons__named_grid__scoped_attribute_count`¶

domain: icons
scene_id: named_grid
task: scoped_attribute_count
module: src/trace_tasks/tasks/icons/named_grid/scoped_attribute_count.py

Program Contract¶

Program: count.scoped_attribute(scene=named_grid, scope=numbered_row_or_column, attribute=shape, output=count)

Candidate set: the visible icon instances, icon attributes, fields, grids, paths, panels, reference items, and labeled option cards inside the numbered_row_or_column objective scope. Operands: visible scene state and prompt-bound operands named by named_grid, numbered_row_or_column, attribute, shape. Operation: evaluate count.scoped_attribute 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. Annotation schema: bbox_set. Query ids: single.

Reasoning Operations¶

Families: filtering, counting

Query IDs¶

row_shape_count
column_shape_count

Generation¶

Default answer support is 1..5. Grid sizes are sampled from every row/column combination in 4..6:

4 x 4
4 x 5
4 x 6
5 x 4
5 x 5
5 x 6
6 x 4
6 x 5
6 x 6

The selected grid size must have enough cells in the queried row or column to support the sampled answer. The queried line contains exactly the sampled answer count of the target shape. Additional target-shape icons are placed outside the queried row/column as distractors so the task requires row/column localization rather than whole-image counting.

Fill style and color are rendered as non-semantic visual variation.

Trace¶

The trace records: - every named icon with row/column indices, row/column numbers, cell bbox, and icon bbox, - the target shape id/name, - the queried axis and one-based row/column number, - the counted cells and off-line target distractor cells, - render-style metadata for the panel title and row/column number-label text legibility, - query, answer, grid-size, line-index, shape, and fill-style probability metadata.