task_puzzles__balance_scale__query_side_relation_label

Program Contract

Program: select_option(balance_query_relation, target=left|right|balanced|not_determined, unknowns=A|B|C, panels=3); scene=balance_scale; scope=query_side_relation_label

Candidate set: the visible balance-scale panels, object symbols, object counts, side relations, and query markers inside the query_side_relation_label objective scope. Operands: visible scene state and prompt-bound operands named by balance_query_relation, left, right, balanced, not_determined, unknowns, A, B, C, panels, balance_scale, query_side_relation_label. Operation: evaluate select_option 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 string schema; generation binds a unique final answer. Annotation witnesses: annotation uses the bbox schema; scalar bbox marks the selected relation option. Query ids: single.

Reasoning Operations

Families: formula_evaluation, matching

2) Scene + task contract

1. Entities/relations: Three balanced pan-scale reference panels over three unknown object labels, plus a query row comparing two symbolic pan expressions and four visual relation options. The internal not_determined relation is displayed to users as Cannot determine.
2. Supported query_id values: single
3. answer_gt.type: string
4. Annotation schema: bbox
5. Alternate annotation forms: none
6. Annotation witness policy: scalar bbox marks the selected relation option.
7. Overlap/touch policy: option bbox covers the selected option card, including its option label and relation text.

3) Prompt contract

1. prompt_bundle_id: puzzles_balance_scale_v1
2. scene_key: balance_scale
3. task_key: balance_scale_query
4. Optional query-id prompt mapping: public single uses prompt query key query_side_relation_label.
5. Required slots:
6. answer output: answer_hint, json_example_answer_only
7. answer+annotation mode: annotation_hint, answer_hint, json_example
8. JSON example validity rule: the example must use one selected-option bbox and one option-label answer.
9. Variant counts: 5 scene templates, 5 query templates, and 5 output templates per output mode.
10. Output modes: answer_only, answer_and_annotation

4) Determinism + constraints

1. Seed namespaces used: task-local relation construction/object-order namespaces plus shared balance scene/style/font/unit-size namespaces.
2. Unique-answer policy: left, right, and balanced targets must be uniquely implied by the reference scales over the configured weight support; not_determined targets, displayed as Cannot determine, must allow more than one relation over that support.
3. Answer-balance policy: target relation is sampled uniformly from left, right, balanced, and not_determined using the task RNG; the seed only makes that equal-weight draw reproducible.
4. Reject/resample conditions: identical pan expressions, missing query relation bbox, non-four-option query row, determined targets with ambiguous relation, or not_determined targets with only one possible relation raise and retry within max_attempts.
5. No-auto-relaxation guarantee: semantic constraints are not relaxed; generation retries rather than accepting ambiguous determined comparisons or determined unanswerable cases.

5) Tests

1. Determinism test: tests/test_puzzles_balance_scale_tasks.py::test_query_side_relation_task_is_deterministic
2. Answer/annotation consistency test: tests/test_puzzles_balance_scale_tasks.py::test_query_side_relation_task_emits_public_contract
3. Prompt metadata/placeholder test: covered by source-layout checks and prompt-system tests.
4. Constraint-specific tests: tests/test_puzzles_balance_scale_tasks.py::test_query_side_relation_targets_are_valid_and_balanced