diff --git a/README.md b/README.md
index bb0f796..8b1db21 100644
--- a/README.md
+++ b/README.md
@@ -51,14 +51,14 @@ try:
         model="gpt-5",
         messages=[{"role": "user", "content": "Hello world"}],
     )
-    print(chat.llm_response.choices[0].message.content)
+    print(chat.choices[0].message.content)
 
     # Or with the Responses API
     resp = client.responses.create(
         model="gpt-5",
         input="What are the main features of your premium plan?",
     )
-    print(resp.llm_response.output_text)
+    print(resp.output_text)
 except GuardrailTripwireTriggered as e:
     print(f"Guardrail triggered: {e}")
 ```
diff --git a/docs/index.md b/docs/index.md
index f4239e0..4640aaa 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -35,7 +35,7 @@ response = await client.responses.create(
     input="Hello"
 )
 # Guardrails run automatically
-print(response.llm_response.output_text)
+print(response.output_text)
 ```
 
 ## Next Steps
diff --git a/docs/quickstart.md b/docs/quickstart.md
index fe91f01..5c4695d 100644
--- a/docs/quickstart.md
+++ b/docs/quickstart.md
@@ -70,8 +70,8 @@ async def main():
             input="Hello world"
         )
         
-        # Access OpenAI response via .llm_response
-        print(response.llm_response.output_text)
+        # Access OpenAI response attributes directly
+        print(response.output_text)
         
     except GuardrailTripwireTriggered as exc:
         print(f"Guardrail triggered: {exc.guardrail_result.info}")
@@ -79,7 +79,7 @@ async def main():
 asyncio.run(main())
 ```
 
-**That's it!** Your existing OpenAI code now includes automatic guardrail validation based on your pipeline configuration. Just use `response.llm_response` instead of `response`.
+**That's it!** Your existing OpenAI code now includes automatic guardrail validation based on your pipeline configuration. The response object acts as a drop-in replacement for OpenAI responses with added guardrail results.
 
 ## Multi-Turn Conversations
 
@@ -98,7 +98,7 @@ while True:
             model="gpt-4o"
         )
         
-        response_content = response.llm_response.choices[0].message.content
+        response_content = response.choices[0].message.content
         print(f"Assistant: {response_content}")
         
         # ✅ Only append AFTER guardrails pass
diff --git a/docs/ref/checks/hallucination_detection.md b/docs/ref/checks/hallucination_detection.md
index ffc2043..0616902 100644
--- a/docs/ref/checks/hallucination_detection.md
+++ b/docs/ref/checks/hallucination_detection.md
@@ -76,7 +76,7 @@ response = await client.responses.create(
 )
 
 # Guardrails automatically validate against your reference documents
-print(response.llm_response.output_text)
+print(response.output_text)
 ```
 
 ### How It Works
diff --git a/docs/tripwires.md b/docs/tripwires.md
index 89cb6b2..5b261cd 100644
--- a/docs/tripwires.md
+++ b/docs/tripwires.md
@@ -25,7 +25,7 @@ try:
         model="gpt-5",
         input="Tell me a secret"
     )
-    print(response.llm_response.output_text)
+    print(response.output_text)
     
 except GuardrailTripwireTriggered as exc:
     print(f"Guardrail triggered: {exc.guardrail_result.info}")
diff --git a/examples/basic/azure_implementation.py b/examples/basic/azure_implementation.py
index c475103..4279e25 100644
--- a/examples/basic/azure_implementation.py
+++ b/examples/basic/azure_implementation.py
@@ -75,7 +75,7 @@ async def process_input(
         )
 
         # Extract the response content from the GuardrailsResponse
-        response_text = response.llm_response.choices[0].message.content
+        response_text = response.choices[0].message.content
 
         # Only show output if all guardrails pass
         print(f"\nAssistant: {response_text}")
diff --git a/examples/basic/hello_world.py b/examples/basic/hello_world.py
index da53e7f..3509753 100644
--- a/examples/basic/hello_world.py
+++ b/examples/basic/hello_world.py
@@ -48,14 +48,12 @@ async def process_input(
             model="gpt-4.1-mini",
             previous_response_id=response_id,
         )
-
-        console.print(f"\nAssistant output: {response.llm_response.output_text}", end="\n\n")
-
+        console.print(f"\nAssistant output: {response.output_text}", end="\n\n")
         # Show guardrail results if any were run
         if response.guardrail_results.all_results:
             console.print(f"[dim]Guardrails checked: {len(response.guardrail_results.all_results)}[/dim]")
 
-        return response.llm_response.id
+        return response.id
 
     except GuardrailTripwireTriggered:
         raise
diff --git a/examples/basic/local_model.py b/examples/basic/local_model.py
index a3d5c2f..7aea228 100644
--- a/examples/basic/local_model.py
+++ b/examples/basic/local_model.py
@@ -48,7 +48,7 @@ async def process_input(
         )
 
         # Access response content using standard OpenAI API
-        response_content = response.llm_response.choices[0].message.content
+        response_content = response.choices[0].message.content
         console.print(f"\nAssistant output: {response_content}", end="\n\n")
 
         # Add to conversation history
diff --git a/examples/basic/multi_bundle.py b/examples/basic/multi_bundle.py
index 4bdac20..a7cd4fe 100644
--- a/examples/basic/multi_bundle.py
+++ b/examples/basic/multi_bundle.py
@@ -66,15 +66,15 @@ async def process_input(
     with Live(output_text, console=console, refresh_per_second=10) as live:
         try:
             async for chunk in stream:
-                # Access streaming response exactly like native OpenAI API through .llm_response
-                if hasattr(chunk.llm_response, "delta") and chunk.llm_response.delta:
-                    output_text += chunk.llm_response.delta
+                # Access streaming response exactly like native OpenAI API (flattened)
+                if hasattr(chunk, "delta") and chunk.delta:
+                    output_text += chunk.delta
                     live.update(output_text)
 
             # Get the response ID from the final chunk
             response_id_to_return = None
-            if hasattr(chunk.llm_response, "response") and hasattr(chunk.llm_response.response, "id"):
-                response_id_to_return = chunk.llm_response.response.id
+            if hasattr(chunk, "response") and hasattr(chunk.response, "id"):
+                response_id_to_return = chunk.response.id
 
             return response_id_to_return
 
diff --git a/examples/basic/multiturn_chat_with_alignment.py b/examples/basic/multiturn_chat_with_alignment.py
index 4ff9af2..581bb59 100644
--- a/examples/basic/multiturn_chat_with_alignment.py
+++ b/examples/basic/multiturn_chat_with_alignment.py
@@ -235,7 +235,7 @@ async def main(malicious: bool = False) -> None:
                     tools=tools,
                 )
                 print_guardrail_results("initial", resp)
-                choice = resp.llm_response.choices[0]
+                choice = resp.choices[0]
                 message = choice.message
                 tool_calls = getattr(message, "tool_calls", []) or []
 
@@ -327,7 +327,7 @@ async def main(malicious: bool = False) -> None:
                     )
 
                     print_guardrail_results("final", resp)
-                    final_message = resp.llm_response.choices[0].message
+                    final_message = resp.choices[0].message
                     console.print(
                         Panel(
                             final_message.content or "(no output)",
diff --git a/examples/basic/pii_mask_example.py b/examples/basic/pii_mask_example.py
index 5d4dd4b..abcf5dd 100644
--- a/examples/basic/pii_mask_example.py
+++ b/examples/basic/pii_mask_example.py
@@ -90,7 +90,7 @@ async def process_input(
         )
 
         # Show the LLM response (already masked if PII was detected)
-        content = response.llm_response.choices[0].message.content
+        content = response.choices[0].message.content
         console.print(f"\n[bold blue]Assistant output:[/bold blue] {content}\n")
 
         # Show PII masking information if detected in pre-flight
diff --git a/examples/basic/structured_outputs_example.py b/examples/basic/structured_outputs_example.py
index 1d2414a..d86e87d 100644
--- a/examples/basic/structured_outputs_example.py
+++ b/examples/basic/structured_outputs_example.py
@@ -56,11 +56,11 @@ async def extract_user_info(
         )
 
         # Access the parsed structured output
-        user_info = response.llm_response.output_parsed
+        user_info = response.output_parsed
         print(f"✅ Successfully extracted: {user_info.name}, {user_info.age}, {user_info.email}")
 
         # Return user info and response ID (only returned if guardrails pass)
-        return user_info, response.llm_response.id
+        return user_info, response.id
 
     except GuardrailTripwireTriggered:
         # Guardrail blocked - no response ID returned, conversation history unchanged
diff --git a/examples/basic/suppress_tripwire.py b/examples/basic/suppress_tripwire.py
index 19f9311..2ffb8d7 100644
--- a/examples/basic/suppress_tripwire.py
+++ b/examples/basic/suppress_tripwire.py
@@ -68,8 +68,8 @@ async def process_input(
         else:
             console.print("[bold green]No guardrails triggered.[/bold green]")
 
-        console.print(f"\n[bold blue]Assistant output:[/bold blue] {response.llm_response.output_text}\n")
-        return response.llm_response.id
+        console.print(f"\n[bold blue]Assistant output:[/bold blue] {response.output_text}\n")
+        return response.id
 
     except Exception as e:
         console.print(f"[bold red]Error: {e}[/bold red]")
diff --git a/examples/hallucination_detection/run_hallucination_detection.py b/examples/hallucination_detection/run_hallucination_detection.py
index f65ecb2..f901cf4 100644
--- a/examples/hallucination_detection/run_hallucination_detection.py
+++ b/examples/hallucination_detection/run_hallucination_detection.py
@@ -52,7 +52,7 @@ async def main():
                 model="gpt-4.1-mini",
             )
 
-            response_content = response.llm_response.choices[0].message.content
+            response_content = response.choices[0].message.content
             console.print(
                 Panel(
                     f"[bold green]Tripwire not triggered[/bold green]\n\nResponse: {response_content}",
diff --git a/examples/implementation_code/blocking/blocking_completions.py b/examples/implementation_code/blocking/blocking_completions.py
index ef21fb1..7a57fd0 100644
--- a/examples/implementation_code/blocking/blocking_completions.py
+++ b/examples/implementation_code/blocking/blocking_completions.py
@@ -25,7 +25,7 @@ async def process_input(
             model="gpt-4.1-mini",
         )
 
-        response_content = response.llm_response.choices[0].message.content
+        response_content = response.choices[0].message.content
         print(f"\nAssistant: {response_content}")
 
         # Guardrails passed - now safe to add to conversation history
diff --git a/examples/implementation_code/blocking/blocking_responses.py b/examples/implementation_code/blocking/blocking_responses.py
index 1209764..e442a66 100644
--- a/examples/implementation_code/blocking/blocking_responses.py
+++ b/examples/implementation_code/blocking/blocking_responses.py
@@ -18,9 +18,9 @@ async def process_input(guardrails_client: GuardrailsAsyncOpenAI, user_input: st
         # including pre-flight, input, and output stages, plus the LLM call
         response = await guardrails_client.responses.create(input=user_input, model="gpt-4.1-mini", previous_response_id=response_id)
 
-        print(f"\nAssistant: {response.llm_response.output_text}")
+        print(f"\nAssistant: {response.output_text}")
 
-        return response.llm_response.id
+        return response.id
 
     except GuardrailTripwireTriggered:
         # GuardrailsClient automatically handles tripwire exceptions
diff --git a/examples/implementation_code/streaming/streaming_completions.py b/examples/implementation_code/streaming/streaming_completions.py
index 2af0a09..6c62776 100644
--- a/examples/implementation_code/streaming/streaming_completions.py
+++ b/examples/implementation_code/streaming/streaming_completions.py
@@ -30,8 +30,8 @@ async def process_input(
         # Stream with output guardrail checks and accumulate response
         response_content = ""
         async for chunk in stream:
-            if chunk.llm_response.choices[0].delta.content:
-                delta = chunk.llm_response.choices[0].delta.content
+            if chunk.choices[0].delta.content:
+                delta = chunk.choices[0].delta.content
                 print(delta, end="", flush=True)
                 response_content += delta
 
diff --git a/examples/implementation_code/streaming/streaming_responses.py b/examples/implementation_code/streaming/streaming_responses.py
index e784906..3bfeb18 100644
--- a/examples/implementation_code/streaming/streaming_responses.py
+++ b/examples/implementation_code/streaming/streaming_responses.py
@@ -26,15 +26,15 @@ async def process_input(guardrails_client: GuardrailsAsyncOpenAI, user_input: st
 
         # Stream with output guardrail checks
         async for chunk in stream:
-            # Access streaming response exactly like native OpenAI API through .llm_response
+            # Access streaming response exactly like native OpenAI API
             # For responses API streaming, check for delta content
-            if hasattr(chunk.llm_response, "delta") and chunk.llm_response.delta:
-                print(chunk.llm_response.delta, end="", flush=True)
+            if hasattr(chunk, "delta") and chunk.delta:
+                print(chunk.delta, end="", flush=True)
 
         # Get the response ID from the final chunk
         response_id_to_return = None
-        if hasattr(chunk.llm_response, "response") and hasattr(chunk.llm_response.response, "id"):
-            response_id_to_return = chunk.llm_response.response.id
+        if hasattr(chunk, "response") and hasattr(chunk.response, "id"):
+            response_id_to_return = chunk.response.id
 
         return response_id_to_return
 
diff --git a/examples/internal_examples/custom_context.py b/examples/internal_examples/custom_context.py
index 511d327..c26e509 100644
--- a/examples/internal_examples/custom_context.py
+++ b/examples/internal_examples/custom_context.py
@@ -58,7 +58,7 @@ async def main() -> None:
                     model="gpt-4.1-nano",
                     messages=messages + [{"role": "user", "content": user_input}],
                 )
-                response_content = response.llm_response.choices[0].message.content
+                response_content = response.choices[0].message.content
                 print("Assistant:", response_content)
 
                 # Guardrails passed - now safe to add to conversation history
diff --git a/src/guardrails/_base_client.py b/src/guardrails/_base_client.py
index c4bb399..b15771b 100644
--- a/src/guardrails/_base_client.py
+++ b/src/guardrails/_base_client.py
@@ -7,9 +7,11 @@
 from __future__ import annotations
 
 import logging
+import warnings
 from dataclasses import dataclass
 from pathlib import Path
 from typing import Any, Final, Union
+from weakref import WeakValueDictionary
 
 from openai.types import Completion
 from openai.types.chat import ChatCompletion, ChatCompletionChunk
@@ -23,6 +25,28 @@
 
 logger = logging.getLogger(__name__)
 
+# Track instances that have emitted deprecation warnings
+_warned_instance_ids: WeakValueDictionary[int, Any] = WeakValueDictionary()
+
+
+def _warn_llm_response_deprecation(instance: Any) -> None:
+    """Emit deprecation warning for llm_response access.
+
+    Args:
+        instance: The GuardrailsResponse instance.
+    """
+    instance_id = id(instance)
+    if instance_id not in _warned_instance_ids:
+        warnings.warn(
+            "Accessing 'llm_response' is deprecated. "
+            "Access response attributes directly instead (e.g., use 'response.output_text' "
+            "instead of 'response.llm_response.output_text'). "
+            "The 'llm_response' attribute will be removed in future versions.",
+            DeprecationWarning,
+            stacklevel=3,
+        )
+        _warned_instance_ids[instance_id] = instance
+
 # Type alias for OpenAI response types
 OpenAIResponseType = Union[Completion, ChatCompletion, ChatCompletionChunk, Response]  # noqa: UP007
 
@@ -54,22 +78,91 @@ def triggered_results(self) -> list[GuardrailResult]:
         return [r for r in self.all_results if r.tripwire_triggered]
 
 
-@dataclass(frozen=True, slots=True)
+@dataclass(frozen=True, slots=True, weakref_slot=True)
 class GuardrailsResponse:
-    """Wrapper around any OpenAI response with guardrail results.
+    """OpenAI response with guardrail results.
 
-    This class provides the same interface as OpenAI responses, with additional
-    guardrail results accessible via the guardrail_results attribute.
+    Access OpenAI response attributes directly:
+        response.output_text
+        response.choices[0].message.content
 
-    Users should access content the same way as with OpenAI responses:
-    - For chat completions: response.choices[0].message.content
-    - For responses: response.output_text
-    - For streaming: response.choices[0].delta.content
+    Access guardrail results:
+        response.guardrail_results.preflight
+        response.guardrail_results.input
+        response.guardrail_results.output
     """
 
-    llm_response: OpenAIResponseType  # OpenAI response object (chat completion, response, etc.)
+    _llm_response: OpenAIResponseType
     guardrail_results: GuardrailResults
 
+    def __init__(
+        self,
+        llm_response: OpenAIResponseType | None = None,
+        guardrail_results: GuardrailResults | None = None,
+        *,
+        _llm_response: OpenAIResponseType | None = None,
+    ) -> None:
+        """Initialize GuardrailsResponse.
+
+        Args:
+            llm_response: OpenAI response object.
+            guardrail_results: Guardrail results.
+            _llm_response: OpenAI response object (keyword-only alias).
+
+        Raises:
+            TypeError: If arguments are invalid.
+        """
+        if llm_response is not None and _llm_response is not None:
+            msg = "Cannot specify both 'llm_response' and '_llm_response'"
+            raise TypeError(msg)
+
+        if llm_response is None and _llm_response is None:
+            msg = "Must specify either 'llm_response' or '_llm_response'"
+            raise TypeError(msg)
+
+        if guardrail_results is None:
+            msg = "Missing required argument: 'guardrail_results'"
+            raise TypeError(msg)
+
+        response_obj = llm_response if llm_response is not None else _llm_response
+
+        object.__setattr__(self, "_llm_response", response_obj)
+        object.__setattr__(self, "guardrail_results", guardrail_results)
+
+    @property
+    def llm_response(self) -> OpenAIResponseType:
+        """Access underlying OpenAI response (deprecated).
+
+        Returns:
+            OpenAI response object.
+        """
+        _warn_llm_response_deprecation(self)
+        return self._llm_response
+
+    def __getattr__(self, name: str) -> Any:
+        """Delegate attribute access to underlying OpenAI response.
+
+        Args:
+            name: Attribute name.
+
+        Returns:
+            Attribute value from OpenAI response.
+
+        Raises:
+            AttributeError: If attribute doesn't exist.
+        """
+        return getattr(self._llm_response, name)
+
+    def __dir__(self) -> list[str]:
+        """List all available attributes including delegated ones.
+
+        Returns:
+            Sorted list of attribute names.
+        """
+        own_attrs = set(object.__dir__(self))
+        delegated_attrs = set(dir(self._llm_response))
+        return sorted(own_attrs | delegated_attrs)
+
 
 class GuardrailsBaseClient:
     """Base class with shared functionality for guardrails clients."""
@@ -135,7 +228,7 @@ def _create_guardrails_response(
             output=output_results,
         )
         return GuardrailsResponse(
-            llm_response=llm_response,
+            _llm_response=llm_response,
             guardrail_results=guardrail_results,
         )
 
diff --git a/tests/unit/test_response_flattening.py b/tests/unit/test_response_flattening.py
new file mode 100644
index 0000000..9597043
--- /dev/null
+++ b/tests/unit/test_response_flattening.py
@@ -0,0 +1,415 @@
+"""Tests for GuardrailsResponse attribute delegation and deprecation warnings."""
+
+from __future__ import annotations
+
+import warnings
+from types import SimpleNamespace
+from typing import Any
+
+import pytest
+
+from guardrails._base_client import GuardrailResults, GuardrailsResponse
+from guardrails.types import GuardrailResult
+
+
+def _create_mock_chat_completion() -> Any:
+    """Create a mock ChatCompletion response."""
+    return SimpleNamespace(
+        id="chatcmpl-123",
+        choices=[
+            SimpleNamespace(
+                index=0,
+                message=SimpleNamespace(content="Hello, world!", role="assistant"),
+                finish_reason="stop",
+            )
+        ],
+        model="gpt-4",
+        usage=SimpleNamespace(prompt_tokens=10, completion_tokens=5, total_tokens=15),
+    )
+
+
+def _create_mock_response() -> Any:
+    """Create a mock Response (Responses API) response."""
+    return SimpleNamespace(
+        id="resp-123",
+        output_text="Hello from responses API!",
+        conversation=SimpleNamespace(id="conv-123"),
+    )
+
+
+def _create_mock_guardrail_results() -> GuardrailResults:
+    """Create mock guardrail results."""
+    return GuardrailResults(
+        preflight=[GuardrailResult(tripwire_triggered=False, info={"stage": "preflight"})],
+        input=[GuardrailResult(tripwire_triggered=False, info={"stage": "input"})],
+        output=[GuardrailResult(tripwire_triggered=False, info={"stage": "output"})],
+    )
+
+
+def test_direct_attribute_access_works() -> None:
+    """Test that attributes can be accessed directly without llm_response."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        assert response.id == "chatcmpl-123"  # noqa: S101
+        assert response.model == "gpt-4"  # noqa: S101
+        assert response.choices[0].message.content == "Hello, world!"  # noqa: S101
+        assert response.usage.total_tokens == 15  # noqa: S101
+
+
+def test_responses_api_direct_access_works() -> None:
+    """Test that Responses API attributes can be accessed directly."""
+    mock_llm_response = _create_mock_response()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        assert response.id == "resp-123"  # noqa: S101
+        assert response.output_text == "Hello from responses API!"  # noqa: S101
+        assert response.conversation.id == "conv-123"  # noqa: S101
+
+
+def test_guardrail_results_access_no_warning() -> None:
+    """Test that accessing guardrail_results does NOT emit deprecation warning."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        assert response.guardrail_results is not None  # noqa: S101
+        assert len(response.guardrail_results.preflight) == 1  # noqa: S101
+        assert len(response.guardrail_results.input) == 1  # noqa: S101
+        assert len(response.guardrail_results.output) == 1  # noqa: S101
+
+
+def test_llm_response_access_emits_deprecation_warning() -> None:
+    """Test that accessing llm_response emits a deprecation warning."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    with pytest.warns(DeprecationWarning, match="Accessing 'llm_response' is deprecated"):
+        _ = response.llm_response
+
+
+def test_llm_response_chained_access_emits_warning() -> None:
+    """Test that accessing llm_response.attribute emits warning (only once)."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    with pytest.warns(DeprecationWarning, match="Accessing 'llm_response' is deprecated"):
+        _ = response.llm_response.id
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        _ = response.llm_response.model  # Should not raise
+
+
+def test_hasattr_works_correctly() -> None:
+    """Test that hasattr works correctly for delegated attributes."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        assert hasattr(response, "id")  # noqa: S101
+        assert hasattr(response, "choices")  # noqa: S101
+        assert hasattr(response, "model")  # noqa: S101
+        assert hasattr(response, "guardrail_results")  # noqa: S101
+        assert not hasattr(response, "nonexistent_attribute")  # noqa: S101
+
+
+def test_getattr_works_correctly() -> None:
+    """Test that getattr works correctly for delegated attributes."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        assert response.id == "chatcmpl-123"  # noqa: S101
+        assert response.model == "gpt-4"  # noqa: S101
+        assert getattr(response, "nonexistent", "default") == "default"  # noqa: S101
+
+
+def test_attribute_error_for_missing_attributes() -> None:
+    """Test that AttributeError is raised for missing attributes."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    with pytest.raises(AttributeError):
+        _ = response.nonexistent_attribute
+
+
+def test_method_calls_work() -> None:
+    """Test that method calls on delegated objects work correctly."""
+    mock_llm_response = SimpleNamespace(
+        id="resp-123",
+        custom_method=lambda: "method result",
+    )
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        assert response.custom_method() == "method result"  # noqa: S101
+
+
+def test_nested_attribute_access_works() -> None:
+    """Test that nested attribute access works correctly."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    # Nested access should work without warnings
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        assert response.choices[0].message.content == "Hello, world!"  # noqa: S101
+        assert response.choices[0].message.role == "assistant"  # noqa: S101
+        assert response.choices[0].finish_reason == "stop"  # noqa: S101
+
+
+def test_property_access_works() -> None:
+    """Test that property access on delegated objects works correctly."""
+    # Create a mock with a property
+    class MockResponse:
+        @property
+        def computed_value(self) -> str:
+            return "computed"
+
+    mock_llm_response = MockResponse()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    # Property access should work without warnings
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        assert response.computed_value == "computed"  # noqa: S101
+
+
+def test_backward_compatibility_still_works() -> None:
+    """Test that old pattern (response.llm_response.attr) still works despite warning."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    # Old pattern should still work (with warning on first access)
+    with pytest.warns(DeprecationWarning):
+        assert response.llm_response.id == "chatcmpl-123"  # noqa: S101
+
+    # Subsequent accesses should work without warnings
+    with warnings.catch_warnings():
+        warnings.simplefilter("error")
+        assert response.llm_response.model == "gpt-4"  # noqa: S101
+        assert response.llm_response.choices[0].message.content == "Hello, world!"  # noqa: S101
+
+
+def test_deprecation_warning_message_content() -> None:
+    """Test that the deprecation warning contains the expected message."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    # Check the full warning message
+    with warnings.catch_warnings(record=True) as w:
+        warnings.simplefilter("always")
+        _ = response.llm_response
+
+        assert len(w) == 1  # noqa: S101
+        assert issubclass(w[0].category, DeprecationWarning)  # noqa: S101
+        assert "Accessing 'llm_response' is deprecated" in str(w[0].message)  # noqa: S101
+        assert "response.output_text" in str(w[0].message)  # noqa: S101
+        assert "future versions" in str(w[0].message)  # noqa: S101
+
+
+def test_warning_only_once_per_instance() -> None:
+    """Test that deprecation warning is only emitted once per instance."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    # Track all warnings
+    with warnings.catch_warnings(record=True) as w:
+        warnings.simplefilter("always")
+
+        # Access llm_response multiple times (simulating streaming chunks)
+        _ = response.llm_response
+        _ = response.llm_response.id
+        _ = response.llm_response.model
+        _ = response.llm_response.choices
+
+        # Should only have ONE warning despite multiple accesses
+        deprecation_warnings = [warning for warning in w if issubclass(warning.category, DeprecationWarning)]
+        assert len(deprecation_warnings) == 1  # noqa: S101
+
+
+def test_separate_instances_warn_independently() -> None:
+    """Test that different GuardrailsResponse instances warn independently."""
+    mock_llm_response1 = _create_mock_chat_completion()
+    mock_llm_response2 = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response1 = GuardrailsResponse(
+        _llm_response=mock_llm_response1,
+        guardrail_results=guardrail_results,
+    )
+
+    response2 = GuardrailsResponse(
+        _llm_response=mock_llm_response2,
+        guardrail_results=guardrail_results,
+    )
+
+    # Track all warnings
+    with warnings.catch_warnings(record=True) as w:
+        warnings.simplefilter("always")
+
+        # Each instance should warn once
+        _ = response1.llm_response
+        _ = response2.llm_response
+
+        # Multiple accesses to same instance should not warn again
+        _ = response1.llm_response
+        _ = response2.llm_response
+
+        # Should have exactly TWO warnings (one per instance)
+        deprecation_warnings = [warning for warning in w if issubclass(warning.category, DeprecationWarning)]
+        assert len(deprecation_warnings) == 2  # noqa: S101
+
+
+def test_init_backward_compatibility_with_llm_response_param() -> None:
+    """Test that __init__ accepts both llm_response and _llm_response parameters."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    # Positional arguments (original order) should work
+    response_positional = GuardrailsResponse(mock_llm_response, guardrail_results)
+    assert response_positional.id == "chatcmpl-123"  # noqa: S101
+    assert response_positional.guardrail_results == guardrail_results  # noqa: S101
+
+    # Old keyword parameter name should work (backward compatibility)
+    response_old = GuardrailsResponse(
+        llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+    assert response_old.id == "chatcmpl-123"  # noqa: S101
+
+    # New keyword parameter name should work (keyword-only)
+    response_new = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+    assert response_new.id == "chatcmpl-123"  # noqa: S101
+
+    # Both llm_response parameters should raise TypeError
+    with pytest.raises(TypeError, match="Cannot specify both"):
+        GuardrailsResponse(
+            llm_response=mock_llm_response,
+            _llm_response=mock_llm_response,
+            guardrail_results=guardrail_results,
+        )
+
+    # Neither llm_response parameter should raise TypeError
+    with pytest.raises(TypeError, match="Must specify either"):
+        GuardrailsResponse(guardrail_results=guardrail_results)
+
+    # Missing guardrail_results should raise TypeError
+    with pytest.raises(TypeError, match="Missing required argument"):
+        GuardrailsResponse(llm_response=mock_llm_response)
+
+
+def test_dir_includes_delegated_attributes() -> None:
+    """Test that dir() includes attributes from the underlying llm_response."""
+    mock_llm_response = _create_mock_chat_completion()
+    guardrail_results = _create_mock_guardrail_results()
+
+    response = GuardrailsResponse(
+        _llm_response=mock_llm_response,
+        guardrail_results=guardrail_results,
+    )
+
+    # Get all attributes via dir()
+    attrs = dir(response)
+
+    # Should include GuardrailsResponse's own attributes
+    assert "guardrail_results" in attrs  # noqa: S101
+    assert "llm_response" in attrs  # noqa: S101
+    assert "_llm_response" in attrs  # noqa: S101
+
+    # Should include delegated attributes from llm_response
+    assert "id" in attrs  # noqa: S101
+    assert "model" in attrs  # noqa: S101
+    assert "choices" in attrs  # noqa: S101
+
+    # Should be sorted
+    assert attrs == sorted(attrs)  # noqa: S101
+
+    # Verify dir() on llm_response and response have overlap
+    llm_attrs = set(dir(mock_llm_response))
+    response_attrs = set(attrs)
+    # All llm_response attributes should be in response's dir()
+    assert llm_attrs.issubset(response_attrs)  # noqa: S101
+