diff --git a/py/src/braintrust/__init__.py b/py/src/braintrust/__init__.py
index 1e66fced..f568fdb0 100644
--- a/py/src/braintrust/__init__.py
+++ b/py/src/braintrust/__init__.py
@@ -50,6 +50,17 @@ def is_equal(expected, output):
 
 # Check env var at import time for auto-instrumentation
 import os
+from typing import TYPE_CHECKING
+
+
+if TYPE_CHECKING:
+    # These names must precede the generated wildcard so type checkers keep the
+    # runtime resource classes for the four names shared by both modules.
+    from .logger import Dataset as Dataset  # noqa: I001
+    from .logger import Experiment as Experiment
+    from .logger import Project as Project
+    from .logger import Prompt as Prompt
+    from .generated_types import *
 
 
 if os.getenv("BRAINTRUST_INSTRUMENT_THREADS", "").lower() in ("true", "1", "yes"):
@@ -60,6 +71,7 @@ def is_equal(expected, output):
     except Exception:
         pass  # Never break on import
 
+from . import generated_types as _generated_types  # noqa: I001
 from .audit import *
 from .auto import auto_instrument as auto_instrument
 from .dataset_pipeline import *
@@ -67,8 +79,14 @@ def is_equal(expected, output):
 from .framework2 import *
 from .functions.invoke import *
 from .functions.stream import *
-from .generated_types import *
-from .integrations.ai_sdk import setup_ai_sdk as setup_ai_sdk
+
+# Keep this before the logger wildcard so its existing runtime collision
+# precedence remains unchanged while new generated names are picked up.
+for _name in _generated_types.__all__:
+    if _name not in {"Dataset", "Experiment", "Project", "Prompt"}:
+        globals()[_name] = getattr(_generated_types, _name)
+
+from .integrations.ai_sdk import setup_ai_sdk as setup_ai_sdk  # noqa: I001
 from .integrations.anthropic import wrap_anthropic as wrap_anthropic
 from .integrations.instructor import wrap_instructor as wrap_instructor
 from .integrations.litellm import wrap_litellm as wrap_litellm
@@ -77,6 +95,10 @@ def is_equal(expected, output):
 from .integrations.pydantic_ai import setup_pydantic_ai as setup_pydantic_ai
 from .logger import *
 from .logger import (
+    Dataset as Dataset,
+    Experiment as Experiment,
+    Project as Project,
+    Prompt as Prompt,
     _internal_get_global_state,  # noqa: F401 # type: ignore[reportUnusedImport]
     _internal_reset_global_state,  # noqa: F401 # type: ignore[reportUnusedImport]
     _internal_with_custom_background_logger,  # noqa: F401 # type: ignore[reportUnusedImport]
diff --git a/py/src/braintrust/type_tests/test_public_exports.py b/py/src/braintrust/type_tests/test_public_exports.py
index 928f6716..0defbbf1 100644
--- a/py/src/braintrust/type_tests/test_public_exports.py
+++ b/py/src/braintrust/type_tests/test_public_exports.py
@@ -1,15 +1,18 @@
-"""Regression test for pyright's ``reportPrivateImportUsage`` on top-level ``braintrust`` symbols.
+"""Regression tests for top-level ``braintrust`` symbols.
 
-Without PEP 484 ``as``-aliasing in ``braintrust/__init__.py``, pyright flags
-``from braintrust import auto_instrument`` (and peers) as private in a
-``py.typed`` consumer. The local ``pyrightconfig.json`` turns the rule into
-an error so this file breaks ``nox -s test_types`` if someone regresses the
-aliasing pattern.
+The static resource check keeps mypy from resolving generated ``TypedDict``
+names instead of the public runtime classes. The runtime checks cover PEP 484
+aliasing for pyright's ``reportPrivateImportUsage`` rule and keep generated
+exports synchronized with the package root.
 """
 
+import subprocess
+import sys
+
 import braintrust
 import pytest
 from braintrust import (
+    Acl,
     auto_instrument,
     setup_ai_sdk,
     setup_pydantic_ai,
@@ -31,7 +34,42 @@
 ]
 
 
+def accepts_public_resource_types(
+    experiment: braintrust.Experiment,
+    dataset: braintrust.Dataset,
+    project: braintrust.Project,
+    prompt: braintrust.Prompt,
+    acl: Acl,
+) -> None:
+    experiment.fetch()
+    dataset.fetch()
+    _ = project.name
+    prompt.build()
+    _ = acl["id"]
+
+
 @pytest.mark.parametrize("name,imported", _PUBLIC_SYMBOLS)
 def test_top_level_public_symbol(name: str, imported: object) -> None:
     assert callable(imported)
     assert callable(getattr(braintrust, name))
+
+
+def test_generated_exports_follow_generated_all() -> None:
+    script = """
+import importlib
+
+import braintrust
+from braintrust import generated_types
+
+future_type = type("FutureGeneratedType", (), {})
+generated_types.FutureGeneratedType = future_type
+generated_types.__all__.append("FutureGeneratedType")
+try:
+    importlib.reload(braintrust)
+    assert braintrust.FutureGeneratedType is future_type
+finally:
+    generated_types.__all__.remove("FutureGeneratedType")
+    del generated_types.FutureGeneratedType
+"""
+
+    subprocess.run([sys.executable, "-c", script], check=True)