task_icons__pair_grid__reference_transform_match_label

Identity

• domain: icons
• scene_id: pair_grid
• module: src/trace_tasks/tasks/icons/pair_grid/reference_transform_match_label.py
• prompt bundle: icons_pair_grid_v1

Program Contract

Program: selection.reference_relation_match(scene=pair_grid, scope=labeled_scene_cells, reference_relation=geometric_transform, output=option_letter)

Candidate set: the visible icon instances, icon attributes, fields, grids, paths, panels, reference items, and labeled option cards inside the labeled_scene_cells objective scope. Operands: visible scene state and prompt-bound operands named by pair_grid, labeled_scene_cells, reference_relation, geometric_transform. Operation: evaluate selection.reference_relation_match 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 option_letter schema; generation binds a unique final answer. Annotation witnesses: annotation uses the bbox schema; the prompt/annotation contract defines the minimal visual witnesses. Query ids: single.

Reasoning Operations

Families: transformation, matching

Contract

Renders a Reference before/after icon pair and a labeled Scene grid of four before/after icon-pair cells, then asks which labeled Scene cell shows the same geometric transformation as the Reference pair. Exactly one Scene cell is correct by construction.

Supported query ids: - single

Answer schema: option_letter. Annotation schema: scalar bbox around the selected Scene cell.

Notes

The geometric transform id is sampled from rot90|rot180|rot270|flip_h|flip_v and recorded as trace metadata, not as a public query id. The task uses asymmetric icons so distractor transforms remain visually distinct from the Reference transform.

Renderer metadata records sampled palette/style, panel-header and cell-label text-legibility metadata, and per-icon noise edits.