task_puzzles__cell_board__shortest_path_length_value¶
Program Contract¶
Program: shortest_path_length(cell_board, start=green_S_cell, goal=red_G_cell, passable=non_wall_cells, adjacency=orthogonal_4_neighbor); scene=cell_board; scope=shortest_path_length_value
Candidate set: the visible grid cells, cell colors/states, labels, walls, start/goal markers, and mirror or connectivity cues inside the shortest_path_length_value objective scope.
Operands: visible scene state and prompt-bound operands named by cell_board, start, green_S_cell, goal, red_G_cell, passable, non_wall_cells, adjacency, orthogonal_4_neighbor, shortest_path_length_value.
Operation: evaluate shortest_path_length over the candidate set using the visible states, constraints, transforms, comparisons, counts, paths, or option-selection rules 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 segment_set schema; one image-pixel segment for each adjacent step along a shortest path from S to G; segment count equals the answer.
Query ids: single.
Reasoning Operations¶
Families: counting, ranking, topology
2) Scene + task contract¶
- Entities/relations: A rectangular board with dark wall cells, light passable cells, a green start cell marked
S, a red goal cell markedG, and disconnected passable distractors. - Supported
query_idvalues:single answer_gt.type:integer- Annotation schema:
segment_set - Alternate annotation forms: none
- Annotation witness policy: one image-pixel segment for each adjacent step along a shortest path from
StoG; segment count equals the answer. - Overlap/touch policy: walls block movement and diagonal touching is not a step.
3) Prompt contract¶
prompt_bundle_id:puzzles_cell_board_v1scene_key:cell_boardtask_key:cell_board_topology_query- Prompt query key:
shortest_path_length_value - Required slots: output-mode slots from the prompt bundle.
- JSON example validity rule: segment-set cardinality equals the integer answer.
- Output modes:
answer_only,answer_and_annotation
4) Determinism + constraints¶
- Seed namespaces used: task-local shortest-path namespaces plus shared cell-board layout/style/font/noise namespaces.
- Unique-answer policy: construction creates a verified detour corridor whose shortest S-to-G path is 4..8 steps and longer than direct Manhattan distance.
- Reject/resample conditions: path construction, distractor placement, or path-length mismatch raises and retries.
- No-auto-relaxation guarantee: semantic constraints are not relaxed.