diff --git a/kaievolve/config.py b/kaievolve/config.py
index fe0f479..dd1f89c 100644
--- a/kaievolve/config.py
+++ b/kaievolve/config.py
@@ -275,6 +275,12 @@ class PromptConfig:
     # migration so it gets enough shots to matter. None => fall back to the
     # database migration_interval.
     research_director_interval: Optional[int] = None
+    # Dedicated model for the director's (meta-reasoning) call. The director is a
+    # strategy/diagnosis task, so it benefits from a strong reasoning model rather
+    # than whatever the run roster samples. None => use the run roster
+    # (config.llm.models). When set, the director uses ONLY this model, inheriting
+    # connection settings (api_base/key/temperature/...) from the run roster.
+    research_director_model: Optional[str] = None
 
     # Strategy clustering + cluster bandit (Phase 3). Clusters programs by the
     # embedding of their HMRD summary into emergent strategies, then biases
diff --git a/kaievolve/research_director.py b/kaievolve/research_director.py
index 7fb1b78..af61f3b 100644
--- a/kaievolve/research_director.py
+++ b/kaievolve/research_director.py
@@ -96,11 +96,24 @@ def __init__(
         self._last_directive = ""
         self._best_at_last: Optional[float] = None
 
+    def _director_models(self):
+        """Models the director calls with. If prompt.research_director_model is
+        set, the director uses ONLY that model (a strong reasoning model for the
+        meta task), cloning connection settings from the run roster; otherwise it
+        uses the run roster (config.llm.models)."""
+        models = self.config.llm.models
+        name = getattr(self.config.prompt, "research_director_model", None)
+        if name and models:
+            import dataclasses
+
+            return [dataclasses.replace(models[0], name=name, weight=1.0)]
+        return models
+
     def _llm(self):
         if self._ensemble is None:
             from kaievolve.llm.ensemble import LLMEnsemble
 
-            self._ensemble = LLMEnsemble(self.config.llm.models)
+            self._ensemble = LLMEnsemble(self._director_models())
         return self._ensemble
 
     @staticmethod
diff --git a/tests/test_research_director.py b/tests/test_research_director.py
index 4a30f5c..18dc99f 100644
--- a/tests/test_research_director.py
+++ b/tests/test_research_director.py
@@ -113,5 +113,32 @@ def get_best_program(self):
             self.assertIsNone(asyncio.run(d.run(Boom(), 20)))
 
 
+class TestDirectorModelSelection(unittest.TestCase):
+    def _cfg(self, director_model):
+        from kaievolve.config import LLMModelConfig
+
+        cfg = Config()
+        cfg.llm.models = [
+            LLMModelConfig(name="fast/a", api_base="https://x/v1", temperature=0.7),
+            LLMModelConfig(name="fast/b"),
+        ]
+        cfg.prompt.research_director_model = director_model
+        return cfg
+
+    def test_dedicated_model_used_and_inherits_connection(self):
+        with TemporaryDirectory() as tmp:
+            d = ResearchDirector(self._cfg("strong/reasoner"), tmp)
+            ms = d._director_models()
+            self.assertEqual([m.name for m in ms], ["strong/reasoner"])  # only the dedicated model
+            self.assertEqual(
+                ms[0].api_base, "https://x/v1"
+            )  # connection settings cloned from roster
+
+    def test_default_falls_back_to_roster(self):
+        with TemporaryDirectory() as tmp:
+            d = ResearchDirector(self._cfg(None), tmp)
+            self.assertEqual([m.name for m in d._director_models()], ["fast/a", "fast/b"])
+
+
 if __name__ == "__main__":
     unittest.main()