diff --git a/tests/eval/consolidation/test_consolidation_eval.py b/tests/eval/consolidation/test_consolidation_eval.py
index 8b4a88d7..c9f85a23 100644
--- a/tests/eval/consolidation/test_consolidation_eval.py
+++ b/tests/eval/consolidation/test_consolidation_eval.py
@@ -21,6 +21,7 @@
     RejectNewDecision,
     UnifyDecision,
 )
+from reflexio.test_support.llm_mock import assert_litellm_unpatched
 from reflexio.test_support.skip_decorators import skip_low_priority
 from tests.eval.consolidation.case import (
     ConsolidationEvalCase,
@@ -519,6 +520,10 @@ def test_live_consolidation_provider_real(tmp_path):  # pragma: no cover - manua
     the fixture. Asserts only pipeline mechanics (every case produced one of
     the four kinds), never exact kinds. Run manually with API keys +
     RUN_LOW_PRIORITY=1."""
+    # This test lives outside ``tests/e2e_tests/``, so the session-wide
+    # ``litellm.completion`` patch is installed for it. Without this it would
+    # grade canned mock text while reporting a real model.
+    assert_litellm_unpatched()
     from reflexio.server.api_endpoints.request_context import RequestContext
     from reflexio.server.llm.litellm_client import LiteLLMClient, LiteLLMConfig
 
@@ -544,6 +549,10 @@ def test_live_consolidation_provider_real(tmp_path):  # pragma: no cover - manua
 @skip_low_priority
 def test_real_judge_smoke():  # pragma: no cover - manual, costs money
     """Smoke test against a real judge model. Run manually with API keys."""
+    # This test lives outside ``tests/e2e_tests/``, so the session-wide
+    # ``litellm.completion`` patch is installed for it. Without this it would
+    # grade canned mock text while reporting a real model.
+    assert_litellm_unpatched()
     from reflexio.server.llm.litellm_client import LiteLLMClient, LiteLLMConfig
 
     client = LiteLLMClient(LiteLLMConfig(model="claude-haiku-4-5"))
diff --git a/tests/eval/extraction/test_extraction_eval.py b/tests/eval/extraction/test_extraction_eval.py
index 5a6481fc..7c1a8df3 100644
--- a/tests/eval/extraction/test_extraction_eval.py
+++ b/tests/eval/extraction/test_extraction_eval.py
@@ -14,6 +14,7 @@
 import pytest
 from pydantic import BaseModel
 
+from reflexio.test_support.llm_mock import assert_litellm_unpatched
 from reflexio.test_support.skip_decorators import skip_low_priority
 from tests.eval.conftest import _load, _load_rubric
 from tests.eval.extraction.providers import make_extraction_provider
@@ -180,6 +181,10 @@ def test_score_golden_case(extraction_case, extraction_judge):
 @skip_low_priority
 def test_real_judge_smoke():  # pragma: no cover - manual, costs money
     """Smoke test against a real judge model. Run manually with API keys."""
+    # This test lives outside ``tests/e2e_tests/``, so the session-wide
+    # ``litellm.completion`` patch is installed for it. Without this it would
+    # grade canned mock text while reporting a real model.
+    assert_litellm_unpatched()
     from reflexio.server.llm.litellm_client import LiteLLMClient, LiteLLMConfig
 
     rubric = _load_rubric("extraction_rubric.yaml")
@@ -298,6 +303,10 @@ def test_live_extraction_provider_real(tmp_path):  # pragma: no cover - manual
     in range), never exact scores. Run manually with API keys +
     RUN_LOW_PRIORITY=1.
     """
+    # This test lives outside ``tests/e2e_tests/``, so the session-wide
+    # ``litellm.completion`` patch is installed for it. Without this it would
+    # grade canned mock text while reporting a real model.
+    assert_litellm_unpatched()
     from reflexio.server.api_endpoints.request_context import RequestContext
     from reflexio.server.llm.litellm_client import LiteLLMClient, LiteLLMConfig
 
diff --git a/tests/eval/scenarios/test_scenario_eval.py b/tests/eval/scenarios/test_scenario_eval.py
index 9cf4e85a..93e11f62 100644
--- a/tests/eval/scenarios/test_scenario_eval.py
+++ b/tests/eval/scenarios/test_scenario_eval.py
@@ -14,6 +14,7 @@
     RejectNewDecision,
     UnifyDecision,
 )
+from reflexio.test_support.llm_mock import assert_litellm_unpatched
 from reflexio.test_support.skip_decorators import skip_low_priority
 from tests.eval.consolidation.judge import ConsolidationVerdict
 from tests.eval.scenarios.book import _next_id, apply_consolidation
@@ -172,6 +173,10 @@ def test_wrong_consolidation_verdict_fails_scenario() -> None:
 
 @skip_low_priority
 def test_scenario_real(tmp_path) -> None:  # pragma: no cover - manual, costs money
+    # This test lives outside ``tests/e2e_tests/``, so the session-wide
+    # ``litellm.completion`` patch is installed for it. Without this it would
+    # grade canned mock text while reporting a real model.
+    assert_litellm_unpatched()
     from reflexio.server.api_endpoints.request_context import RequestContext
     from reflexio.server.llm.litellm_client import LiteLLMClient, LiteLLMConfig
     from tests.eval.consolidation.providers import (