fix(learning-harness): logprobs are on the choice, not the message

OpenAI-shaped providers put `logprobs` beside the message rather than
inside it, so the confidence signal the classifier is meant to abstain on
was always None -- a knob that reads 'never unsure' whatever the model said.

Measured against DeepSeek while checking: it returns the structure but not
the values. Every chosen token comes back at logprob 0.0 and every
alternative at -9999.0, on a deliberately uncertain prompt as much as a
certain one. The field is there; the information is not.

Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_01T7XD7x5KonMdmGRkp63EQG

src/openultrasast/learning/endpoint.py

@@ -189,12 +189,19 @@ def _provider_of(base_url: str) -> Provider:
 
 
 def _message_of(payload: object) -> Mapping[str, object]:
+    """The reply message, carrying the choice's `logprobs` with it.
+
+    OpenAI-shaped providers put `logprobs` on the *choice*, beside the message rather than inside it. Reading only
+    the message meant the confidence signal was always absent, so a knob meant to let the classifier abstain when
+    it is unsure read as "never unsure" whatever the model said."""
     if isinstance(payload, Mapping):
         choices = payload.get("choices")
         if isinstance(choices, Sequence) and choices and isinstance(choices[0], Mapping):
-            message = choices[0].get("message")
+            choice = choices[0]
+            message = choice.get("message")
             if isinstance(message, Mapping):
-                return message
+                logprobs = choice.get("logprobs")
+                return {**message, "logprobs": logprobs} if logprobs is not None else message
     return {}
 
 

tests/test_learning_endpoint.py

@@ -172,3 +172,33 @@ def test_the_underlying_client_forwards_extra_body_and_keeps_usage(monkeypatch:
     assert payload["usage"] == {"prompt_tokens": 7}
     message = provider._extract_message(payload)
     assert message["reasoning_content"] == "because"  # retained for the caller to replay
+
+
+def test_logprobs_are_read_off_the_choice_where_the_provider_puts_them() -> None:
+    """OpenAI-shaped providers put `logprobs` on the choice, not on the message.
+
+    `_message_of` returned only `choices[0].message`, so the confidence signal the classifier is supposed to
+    abstain on was always `None` — a knob that reads as "never unsure" whatever the model said."""
+    from openultrasast.learning.endpoint import _message_of
+
+    payload = {
+        "choices": [
+            {
+                "message": {"content": "yes", "role": "assistant"},
+                "logprobs": {"content": [{"token": "yes", "logprob": -0.5}, {"token": "!", "logprob": -1.5}]},
+            }
+        ]
+    }
+    message = _message_of(payload)
+    assert message["content"] == "yes"
+    from openultrasast.tool_hunter import chat_response_from_message
+
+    assert chat_response_from_message(message).mean_logprob == pytest.approx(-1.0)
+
+
+def test_a_reply_without_logprobs_reports_none_rather_than_a_number() -> None:
+    from openultrasast.learning.endpoint import _message_of
+    from openultrasast.tool_hunter import chat_response_from_message
+
+    message = _message_of({"choices": [{"message": {"content": "yes"}}]})
+    assert chat_response_from_message(message).mean_logprob is None