Fix/2894 inverted score metadata

pyrit/score/true_false/true_false_inverter_scorer.py

@@ -16,6 +16,7 @@
 )
 from pyrit.score.scorer import Scorer
 from pyrit.score.scorer_prompt_validator import ScorerPromptValidator
+from pyrit.score.score_utils import ORIGINAL_FLOAT_VALUE_KEY
 from pyrit.score.true_false.true_false_scorer import TrueFalseScorer
 
 
@@ -128,6 +129,10 @@ def _invert(self, scores: list[Score]) -> list[Score]:
             )
         else:
             inv_score.score_value = str(True) if not inv_score.get_value() else str(False)
+            if inv_score.score_metadata and ORIGINAL_FLOAT_VALUE_KEY in inv_score.score_metadata:
+                inv_score.score_metadata[ORIGINAL_FLOAT_VALUE_KEY] = 1.0 - float(
+                    inv_score.score_metadata[ORIGINAL_FLOAT_VALUE_KEY]
+                )
             inv_score.score_value_description = "Inverted score: " + str(inv_score.score_value_description)
             inv_score.score_rationale = (
                 f"Inverted score from {scorer_type} result: {inv_score.score_value}\n{inv_score.score_rationale}"

pyrit/score/true_false/true_false_score_aggregator.py

@@ -9,6 +9,7 @@
 from pyrit.score.score_utils import (
     combine_metadata_and_categories,
     format_score_for_rationale,
+    ORIGINAL_FLOAT_VALUE_KEY,
 )
 
 BinaryBoolOp = Callable[[bool | None, bool | None], bool | None]
@@ -144,6 +145,10 @@ def aggregator(scores: Iterable[Score]) -> ScoreAggregatorResult:
             undetermined_msg=undetermined_msg,
         )
         metadata, category = combine_metadata_and_categories(scores_list)
+        # A single threshold score retains its precise float for feedback. Once multiple
+        # scorers are combined, no single child's float can represent the aggregate verdict.
+        if len(scores_list) > 1:
+            metadata.pop(ORIGINAL_FLOAT_VALUE_KEY, None)
 
         return ScoreAggregatorResult(
             value=result,

tests/unit/score/test_true_false_inverter_scorer.py

@@ -0,0 +1,38 @@
+# Copyright (c) Microsoft Corporation.
+# Licensed under the MIT license.
+
+from unittest.mock import MagicMock
+
+from pyrit.models import ComponentIdentifier, Score
+from pyrit.score.score_utils import ORIGINAL_FLOAT_VALUE_KEY, normalize_score_to_float
+from pyrit.score.true_false.true_false_inverter_scorer import TrueFalseInverterScorer
+from pyrit.score.true_false.true_false_scorer import TrueFalseScorer
+
+
+def _score(value: bool, original_float: float) -> Score:
+    return Score(
+        score_value=str(value),
+        score_value_description="threshold",
+        score_type="true_false",
+        score_metadata={ORIGINAL_FLOAT_VALUE_KEY: original_float},
+        message_piece_id="test-piece",
+        scorer_class_identifier=ComponentIdentifier(
+            class_name="FloatScaleThresholdScorer",
+            class_module="tests.unit.score",
+        ),
+    )
+
+
+def test_inverter_complements_original_float_metadata() -> None:
+    wrapped = MagicMock(spec=TrueFalseScorer)
+    wrapped.get_identifier.return_value = ComponentIdentifier(
+        class_name="FloatScaleThresholdScorer",
+        class_module="tests.unit.score",
+    )
+    inverter = TrueFalseInverterScorer(scorer=wrapped)
+
+    inverted = inverter._invert([_score(True, 0.73)])[0]
+
+    assert inverted.get_value() is False
+    assert inverted.score_metadata == {ORIGINAL_FLOAT_VALUE_KEY: 0.27}
+    assert normalize_score_to_float(inverted) == 0.27

tests/unit/score/test_true_false_score_aggregator.py

@@ -276,3 +276,27 @@ def test_generator_of_wrong_type_still_raises():
     )
     with pytest.raises(ValueError, match="must be of type 'true_false'"):
         TrueFalseScoreAggregator.OR(s for s in [bad])
+
+
+def test_single_threshold_float_metadata_is_preserved() -> None:
+    from pyrit.score.score_utils import ORIGINAL_FLOAT_VALUE_KEY
+
+    score = _mk_score(True, prr_id="1")
+    score.score_metadata = {ORIGINAL_FLOAT_VALUE_KEY: 0.73}
+
+    res = TrueFalseScoreAggregator.AND([score])
+
+    assert res.metadata == {ORIGINAL_FLOAT_VALUE_KEY: 0.73}
+
+
+def test_composite_drops_ambiguous_threshold_float_metadata() -> None:
+    from pyrit.score.score_utils import ORIGINAL_FLOAT_VALUE_KEY
+
+    first = _mk_score(True, prr_id="1")
+    first.score_metadata = {ORIGINAL_FLOAT_VALUE_KEY: 0.73}
+    second = _mk_score(True, prr_id="1")
+    second.score_metadata = {ORIGINAL_FLOAT_VALUE_KEY: 0.41}
+
+    res = TrueFalseScoreAggregator.AND([first, second])
+
+    assert ORIGINAL_FLOAT_VALUE_KEY not in res.metadata