task_symbolic__braille_cell__braille_word_read_label

1) Identity

1. Domain: symbolic
2. Scene id: braille_cell
3. Scene: braille_cell
4. Task id: task_symbolic__braille_cell__braille_word_read_label
5. Objective: choose the word option encoded by a multi-cell Braille plate.

Program Contract

Program: braille_cell.word_read(scene=braille_cell, scope=source_braille_word_plate_and_four_word_options, output=option_label)

Candidate set: the four visible word option cards labeled A..D. Operands: the raised-dot patterns in each source Braille cell and the candidate lowercase option words. Operation: decode the source Grade 1 Braille word and select the unique option word that matches it. Output binding: answer is the selected option letter. Annotation witnesses: a bbox_map with source_plate and selected_option roles. Query ids: single.

Reasoning Operations

Families: transformation

2) Scene + Task Contract

1. Public branch metadata: query_id
2. Supported query_id: single
3. Internal prompt/query key: braille_word_read_label
4. Supported non-semantic visual axes:
5. scene_variant: clean_card|notebook_card|exam_scan
6. shared symbolic background/noise style
7. answer_gt.type: string
8. Answer schema: string, one capital-letter option label from A..D.
9. annotation_gt.type: bbox_map
10. Annotation schema: bbox_map
11. Scene contract:
12. one source Braille word plate is shown,
13. four visual word options are labeled A..D,
14. every option word has the same length as the source word,
15. all option words share at least the first character, so the first Braille cell alone cannot identify the answer,
16. exactly one option is the lowercase word encoded by the Braille plate,
17. words use Grade 1 uncontracted lowercase Braille letters only,
18. generated words are sampled from trace_tasks.tasks.shared.word_assets,
19. word length support is 3..5.

3) Prompt Contract

1. Bundle: symbolic_braille_cell_v1
2. scene_key: braille_cell
3. task_key: braille_word_read_label
4. Required slots:
5. scene: object_description
6. answer output: json_output_contract_answer_only, answer_hint, json_example_answer_only
7. answer+annotation mode: json_output_contract, annotation_hint, answer_hint, json_example
8. Modes: answer_only, answer_and_annotation
9. Options are visible word cards in the image, not prompt-only choices.

4) Annotation + Trace Contract

1. Prompt-facing annotation is a bbox_map with two semantic roles:
2. source_plate: bbox around the source Braille word plate,
3. selected_option: bbox around the correct visual word option.
4. Distractor option bboxes, individual Braille cells, individual dots, text bboxes, labels, and panel chrome are not prompt-facing annotation.
5. projected_annotation includes bbox_map and pixel_bbox_map.
6. render_map includes:
7. item_bboxes_px
8. dot_centers_px
9. raised_dot_centers_px
10. cell_dot_centers_px
11. execution_trace.braille_metadata records the target word, word length, target cell patterns, correct option label/id, option words, and shared-prefix sampling metadata.

5) Determinism + Tests

1. Deterministic sampling/rendering from instance_seed.
2. Answers and annotation come from the same finalized source/option layout.
3. Behavior/trace/prompt tests: tests/test_symbolic_braille_tasks.py
4. Source-layout contract tests: tests/test_public_source_layout_contracts.py