"""Allow-list of (dataset, example_id) pairs that have a precomputed
model_answer_short attribution available for the small Qwen 3 4B model.

For these examples, Public Mode renders a side-by-side dual heatmap
("vs Ground Truth" + "vs Model Answer (Wrong)"). All other examples render
the existing single-heatmap layout unchanged.

The compact attributions only exist at the (geomean_jointprob, word)
combination on disk, so `has_wrong_answer_view` returns False for any other
scalarizer or feature level.
"""

from typing import Set, Tuple

WRONG_ANSWER_EXAMPLES: Set[Tuple[str, str]] = {
    ("esnli", f"example_{i}") for i in range(1, 11)
} | {
    ("snarks", f"example_{i}") for i in (3, 5, 6, 8, 9, 10)
} | {
    ("fever", f"example_{i}") for i in (2, 7, 8, 9, 10)
} | {
    ("medical_qa", f"example_{i}") for i in (1, 3, 5, 9)
} | {
    ("bbq_disamb", f"example_{i}") for i in (2, 4, 9)
} | {
    ("causal_judgment", "example_1"),
    ("bar_exam", "example_3"),
}

assert len(WRONG_ANSWER_EXAMPLES) == 30, (
    f"WRONG_ANSWER_EXAMPLES expected 30 entries, got {len(WRONG_ANSWER_EXAMPLES)}"
)


def has_wrong_answer_view(
    dataset: str,
    example_id: str,
    scalarizer: str,
    feature_level: str,
) -> bool:
    if scalarizer != "geomean_jointprob" or feature_level != "word":
        return False
    return (dataset, example_id) in WRONG_ANSWER_EXAMPLES