refactor(language): adopt the public participation API

.github/workflows/umi-v2.yml

@@ -11,9 +11,9 @@ jobs:
   coordinated:
     name: UMI v2
     # Update the workflow and runner revision together.
-    uses: KartikP/brainscore-unified/.github/workflows/offline-tests.yml@529237ada8c243cab0ad67cb53abde350342f4ea
+    uses: KartikP/brainscore-unified/.github/workflows/offline-tests.yml@5dc12f6654f5701b5135c7b288a3327ee704384e
     with:
       candidate: language
       candidate_repository: ${{ github.event.pull_request.head.repo.full_name || github.repository }}
       candidate_revision: ${{ github.event.pull_request.head.sha || github.sha }}
-      unified_revision: 529237ada8c243cab0ad67cb53abde350342f4ea
+      unified_revision: 5dc12f6654f5701b5135c7b288a3327ee704384e

README.md

@@ -15,7 +15,7 @@ the `ArtificialSubject` interface can be scored on.
 
 > **UMI integration:** Existing domain plugins remain supported. Use
 > `BrainScoreModel` for extraction and task helpers, or implement
-> `Subject.interact(session)` for a custom session integration. A native
+> `Subject.participate(session)` for a custom session integration. A native
 > `Subject` does not require `process()`. See
 > [docs/UMI_MIGRATION.md](docs/UMI_MIGRATION.md).
 

brainscore_language/__init__.py

@@ -83,7 +83,7 @@ def _run_score(model_identifier: str, benchmark_identifier: str,
         check_memory(model, benchmark)
 
     # A legacy benchmark calls digest_text and friends; it receives the permanent
-    # facade, which runs those calls as sessions through the subject's interact.
+    # facade, which runs those calls as sessions through the subject's participate.
     from brainscore_core.legacy_facade import subject_for_benchmark
     candidate = subject_for_benchmark(model, benchmark)
 

brainscore_language/compat/unified_adapter.py

@@ -14,9 +14,9 @@
 from brainscore_core.session_view import DeclaredInputSession
 from brainscore_core.streaming import StreamEvent
 from brainscore_core.streaming_helpers import (
-    _drain_stream_events,
-    _drive_behavior_session_via_process,
-    _drive_neural_session_via_process,
+    drain_session_inputs,
+    drive_behavior_session,
+    drive_neural_session,
     measurement_settings,
 )
 
@@ -169,7 +169,7 @@ def out_channels(self) -> Set[str]:
             return set(self._declared_out_channels)
         return super().out_channels | {'behavior'}
 
-    def interact(self, session) -> None:
+    def participate(self, session) -> None:
         """Run a session; a legacy-benchmark session replays the exact ``digest_text``.
 
         A session carrying ``legacy_text`` (from the legacy-benchmark facade)
@@ -187,17 +187,17 @@ def interact(self, session) -> None:
         settings = measurement_settings(session)
         if 'legacy_text' not in settings:
             if channels == ('behavior',):
-                _drive_behavior_session_via_process(self, session)
+                drive_behavior_session(self, session)
             else:
-                _drive_neural_session_via_process(self, session)
+                drive_neural_session(self, session)
             return
         if context is not None and context.task_type != 'passive':
             self.start_task(context)
         for channel in channels:
             if channel.startswith('neural:'):
                 self.start_neural_recording(channel.split(':', 1)[1],
                                             settings.get('recording_type', 'fMRI'))
-        _drain_stream_events(session)
+        drain_session_inputs(session)
         result = self._legacy.digest_text(settings['legacy_text'])
         for channel in channels:
             key = 'neural' if channel.startswith('neural:') else 'behavior'

docs/UMI_MIGRATION.md

@@ -6,7 +6,7 @@
 | --- | --- |
 | Score an existing model | `brainscore.score(model_identifier, benchmark_identifier)` |
 | Reuse extraction and task helpers | `BrainScoreModel`, with `process`, task setup and recording methods |
-| Define custom session behavior | `Subject.interact(session)`, with declared input/output channels |
+| Define custom session behavior | `Subject.participate(session)`, with declared input/output channels |
 | Combine recording and interventions | `Experiment` with a compatible protocol and tools |
 
 `BrainScoreModel` is a `Subject` implementation. A native `Subject` does not need `process()` or a region mapping; a benchmark using those methods requires an implementation that provides them.

requirements-v2.txt

@@ -1,2 +1,2 @@
 # Unpublished v2 core, pinned for standalone CI and documentation builds.
-brainscore-core @ git+https://github.com/brain-score/core.git@0fbb419ff0ba461e6b1e217c845e0fc61d8096a1
+brainscore-core @ git+https://github.com/brain-score/core.git@9e421f634699fdd4693fc6bc9e210dfb16109e7f

tests/test_unified_adapter.py

@@ -279,7 +279,7 @@ def test_process_single_key_fallback(self):
         result = adapter.process(['text'])
         assert result is sentinel
 
-    def test_neural_response_interact_matches_legacy_process_exactly(self):
+    def test_neural_response_participate_matches_legacy_process_exactly(self):
         stimuli = _language_stimulus_set()
         expected = _language_neural_assembly()
 
@@ -305,6 +305,23 @@ def test_neural_response_interact_matches_legacy_process_exactly(self):
             'language_system', 'fMRI'
         )
 
+    @pytest.mark.parametrize('channel', ['behavior', 'neural:language_system'])
+    def test_typed_session_refuses_variance_before_any_legacy_call(self, channel):
+        from brainscore_core.streaming import BufferedSession
+        legacy = _make_legacy_model(
+            region_layer_mapping={'language_system': 'language'})
+        session = BufferedSession(
+            stimulus_set=_language_stimulus_set(),
+            task_context=TaskContext(task_type='next_word'),
+            requested_output_channels=(channel,))
+        session.settings = {'require_variance': True}
+        with pytest.raises(NotImplementedError, match='trial-to-trial variance'):
+            LanguageModelAdapter(legacy).participate(session)
+        legacy.start_behavioral_task.assert_not_called()
+        legacy.start_neural_recording.assert_not_called()
+        legacy.digest_text.assert_not_called()
+        assert session.emitted == []
+
 
 class TestLanguageAdapterLegacyMethods:
 
@@ -449,7 +466,7 @@ def test_load_model_wraps_legacy(self):
         # After load_model, identifier was overwritten to string 'test-gpt2'
         assert model.identifier == 'test-gpt2'
 
-    def test_load_model_wrapped_legacy_interact_scores(self):
+    def test_load_model_wrapped_legacy_participate_scores(self):
         import brainscore_language
         expected = _language_neural_assembly()
         legacy = _make_legacy_model(