FIX: reject a threshold aggregator that does not combine scores

pyrit/score/true_false/float_scale_threshold_scorer.py

@@ -61,7 +61,8 @@ def __init__(
             scorer (FloatScaleScorer): The underlying float scale scorer to use.
             threshold (float): A finite threshold in (0, 1]. Scores >= threshold are True, otherwise False.
             float_scale_aggregator (FloatScaleAggregatorFunc): The aggregator function to use for combining
-                multiple float scale scores. Defaults to FloatScaleScoreAggregator.MAX.
+                multiple float scale scores. Defaults to FloatScaleScoreAggregator.MAX. It must combine the
+                scores into exactly one result; a by-category aggregator is rejected when scoring.
 
         Raises:
             ValueError: If the threshold is non-finite or not in (0, 1].
@@ -172,11 +173,24 @@ def _apply_threshold(
 
         Returns:
             list[Score]: A list containing one completed or undetermined true/false score.
+
+        Raises:
+            ValueError: If the configured aggregator does not combine the scores into exactly one result.
         """
         objective = expectation.objective if expectation else None
 
         # The wrapped scorer's non-applicable result returns before aggregation.
         aggregate_results = self._float_scale_aggregator(scores)
+        if len(aggregate_results) != 1:
+            # A by-category aggregator returns one result per category, and the
+            # threshold can only be applied to a single value. Silently taking the
+            # first one would decide the verdict from one category and drop the rest.
+            raise ValueError(
+                f"{self._float_scale_aggregator.__name__} returned {len(aggregate_results)} results. "  # type: ignore[ty:unresolved-attribute]
+                "FloatScaleThresholdScorer requires an aggregator that combines the scores into exactly "
+                "one result, such as FloatScaleScoreAggregator.MAX; a by-category aggregator such as "
+                "FloatScaleScorerByCategory.MAX cannot be thresholded."
+            )
         aggregate_score = aggregate_results[0]
         aggregate_value = aggregate_score.value
         scorer_type = self._scorer.get_identifier().class_name

tests/unit/score/test_float_scale_threshold_scorer.py

@@ -265,6 +265,78 @@ async def test_float_scale_threshold_scorer_attributes_result_to_aggregate_not_f
     assert "azure_severity" not in score.score_metadata
 
 
+async def test_float_scale_threshold_scorer_rejects_aggregator_that_does_not_combine():
+    """
+    A by-category aggregator returns one result per category, and the threshold can only be
+    applied to a single value. It used to be accepted and only the first result was used, so
+    the verdict came from whichever category sorted first and the others were dropped without
+    a log: with Hate at 0.0 and Violence at 0.9 the caller received False.
+    """
+    from pyrit.score.float_scale.float_scale_score_aggregator import FloatScaleScorerByCategory
+
+    memory = MagicMock(MemoryInterface)
+    mock_identifier = ComponentIdentifier(class_name="MockScorer", class_module="test.mock")
+
+    prompt_id = uuid.uuid4()
+    scorer = MagicMock(spec=MessageFloatScaleScorer)
+    scorer._score_nested_async = AsyncMock(
+        return_value=[
+            Score(
+                score_value=value,
+                score_type="float_scale",
+                score_category=[category],
+                score_rationale=f"{category} rationale",
+                score_metadata=None,
+                message_piece_id=prompt_id,
+                score_value_description="",
+                scorer_class_identifier=mock_identifier,
+                id=uuid.uuid4(),
+            )
+            for category, value in (("Hate", "0.0"), ("Violence", "0.9"))
+        ]
+    )
+    scorer.get_identifier = MagicMock(return_value=mock_identifier)
+
+    with patch.object(CentralMemory, "get_memory_instance", return_value=memory):
+        threshold_scorer = FloatScaleThresholdScorer(
+            scorer=scorer,
+            threshold=0.5,
+            float_scale_aggregator=FloatScaleScorerByCategory.MAX,
+        )
+
+        with pytest.raises(RuntimeError, match="returned 2 results") as exc_info:
+            await threshold_scorer.score_text_async(text="mock example")
+
+    # The message has to name the aggregator that cannot be thresholded, so the caller
+    # knows to pick a combining one instead of guessing why no score came back.
+    assert "FloatScaleScorerByCategory.MAX cannot be thresholded" in str(exc_info.value)
+    memory.add_scores_to_memory_async.assert_not_called()
+
+
+async def test_float_scale_threshold_scorer_rejects_aggregator_that_returns_nothing():
+    """An aggregator that returns no result is a configuration error, not an IndexError."""
+
+    def empty_aggregator(scores):
+        return []
+
+    empty_aggregator.__name__ = "empty_aggregator"
+
+    memory = MagicMock(MemoryInterface)
+    scorer = create_mock_float_scorer(0.9)
+
+    with patch.object(CentralMemory, "get_memory_instance", return_value=memory):
+        threshold_scorer = FloatScaleThresholdScorer(
+            scorer=scorer,
+            threshold=0.5,
+            float_scale_aggregator=empty_aggregator,
+        )
+
+        with pytest.raises(RuntimeError, match="empty_aggregator returned 0 results"):
+            await threshold_scorer.score_text_async(text="mock example")
+
+    memory.add_scores_to_memory_async.assert_not_called()
+
+
 async def test_float_scale_threshold_scorer_single_score_attribution_unchanged():
     """A single wrapped score must keep its own category and rationale, as before."""