FIX: Canonicalize RandomTranslationConverter language order for seeded selection

pyrit/converter/random_translation_converter.py

@@ -78,9 +78,12 @@ def __init__(
             default_languages = SeedDataset.from_yaml_file(
                 RandomTranslationConverter._DEFAULT_LANGUAGES_SEED_PROMPT_PATH
             )
-            self.languages = [prompt.value for prompt in default_languages.prompts]
+            selected_languages = [prompt.value for prompt in default_languages.prompts]
         else:
-            self.languages = languages
+            selected_languages = list(languages)
+        # Random selection and identity share one canonical order, so reordering the same pool neither
+        # changes the seeded translations nor the identifier that is supposed to describe them.
+        self.languages = sorted(selected_languages, key=str.casefold)
 
     def _build_identifier(self) -> ComponentIdentifier:
         """
@@ -95,7 +98,7 @@ def _build_identifier(self) -> ComponentIdentifier:
             params={
                 **llm_identifier.params,
                 **word_level_identifier.params,
-                "languages": sorted(self.languages, key=str.casefold),
+                "languages": self.languages,
             },
             converter_target=self._converter_target.get_identifier(),
         )

tests/unit/converter/test_random_translation_converter.py

@@ -7,6 +7,7 @@
 import pytest
 from unit.mocks import get_mock_target_identifier
 
+from pyrit.common.random_context import configure_random_seed, get_configured_random_seed
 from pyrit.converter import (
     RandomTranslationConverter,
     WordPositionSelectionStrategy,
@@ -81,6 +82,25 @@ def test_random_translation_converter_identifier_canonicalizes_language_order(mo
     assert first.get_identifier().hash == reordered.get_identifier().hash
 
 
+async def test_random_translation_converter_canonicalizes_language_order_for_selection(mock_target) -> None:
+    """Same language pool in a different order must produce the same translations under one seed."""
+    original_seed = get_configured_random_seed()
+    configure_random_seed(seed=42)
+    try:
+        first = RandomTranslationConverter(converter_target=mock_target, languages=["French", "Spanish"])
+        reordered = RandomTranslationConverter(converter_target=mock_target, languages=["Spanish", "French"])
+
+        await first.convert_async(prompt="alpha beta gamma delta epsilon")
+        await reordered.convert_async(prompt="alpha beta gamma delta epsilon")
+    finally:
+        configure_random_seed(seed=original_seed)
+
+    assert first.get_identifier().hash == reordered.get_identifier().hash
+    first_message = mock_target.send_prompt_async.call_args_list[0].kwargs["message"]
+    reordered_message = mock_target.send_prompt_async.call_args_list[1].kwargs["message"]
+    assert first_message.message_pieces[0].original_value == reordered_message.message_pieces[0].original_value
+
+
 def test_random_translation_converter_identifier_distinguishes_language_pools(mock_target) -> None:
     first = RandomTranslationConverter(converter_target=mock_target, languages=["French", "Spanish"])
     different = RandomTranslationConverter(converter_target=mock_target, languages=["German", "Japanese"])