diff --git a/packages/sie_gateway/src/http_error.rs b/packages/sie_gateway/src/http_error.rs
index a38bbffcc..c3deece71 100644
--- a/packages/sie_gateway/src/http_error.rs
+++ b/packages/sie_gateway/src/http_error.rs
@@ -144,10 +144,11 @@ pub mod openai_code {
     pub const RATE_LIMIT_EXCEEDED: &str = "rate_limit_exceeded";
     /// OpenAI-compatible scale-from-zero / worker provisioning signal.
     pub const PROVISIONING: &str = "provisioning";
-    /// Worker terminal for a generation that finished nominally but produced
-    /// no visible output text or tool calls (e.g. private reasoning consumed
-    /// the whole budget, #3104/#3136). Terminal and non-retryable: tokens were
-    /// genuinely consumed, so it settles exactly like ``inference_error``.
+    /// Worker terminal for a generation that stopped but produced no visible
+    /// output text or tool calls (e.g. only private reasoning, #3104/#3136).
+    /// A budget that runs out first keeps ``finish_reason: "length"`` instead.
+    /// Terminal and non-retryable: tokens were genuinely consumed, so it
+    /// settles exactly like ``inference_error``.
     pub const EMPTY_MODEL_OUTPUT: &str = "empty_model_output";
     /// Worker terminal for a guard model whose first output positions carry
     /// no complete Yes/No verdict distribution. Terminal and non-retryable;
diff --git a/packages/sie_sdk/README.md b/packages/sie_sdk/README.md
index 648cd0429..4e52346f4 100644
--- a/packages/sie_sdk/README.md
+++ b/packages/sie_sdk/README.md
@@ -247,8 +247,9 @@ different dimension does not change that setting. Its configured threshold
 produces `Yes` (unsafe) or `No` (safe). A missing or invalid verdict returns
 `invalid_guard_verdict`; never treat an error or an empty response as safe.
 Private reasoning is hidden on both input surfaces. If it consumes the entire
-generation budget without usable output, the request fails with
-`empty_model_output`.
+generation budget, the answer is empty with `finish_reason` `"length"`; raise
+`max_tokens` to leave room for the answer. A model that stops without usable
+output fails with `empty_model_output`.
 
 ## Connecting to SIE Serverless or a managed gateway
 
@@ -466,10 +467,11 @@ the dropped items when every submitted item carried an `id`.
 
 Two generation-specific errors are terminal and never retried:
 
-- `empty_model_output` — the generation finished nominally but produced
-  no visible output text (for example, private reasoning consumed the
-  whole token budget). Tokens were genuinely consumed, so the request
-  is not re-run. Surfaces as `ServerError` with
+- `empty_model_output` — the generation stopped but produced no visible
+  output text (for example, the model emitted only private reasoning).
+  Tokens were genuinely consumed, so the request is not re-run. A
+  budget that runs out during reasoning is not this error: the answer
+  is empty with `finish_reason` `"length"`. Surfaces as `ServerError` with
   `code == "empty_model_output"`; on streaming calls it is raised
   mid-stream with the gateway request id attached for correlation.
 - `MODEL_OUTPUT_PARSE_ERROR` — a strict grammar (`grammar.strict` or
diff --git a/packages/sie_server/src/sie_server/adapters/_generation_base.py b/packages/sie_server/src/sie_server/adapters/_generation_base.py
index 4a750c51b..ae19a25da 100644
--- a/packages/sie_server/src/sie_server/adapters/_generation_base.py
+++ b/packages/sie_server/src/sie_server/adapters/_generation_base.py
@@ -704,6 +704,11 @@ def _candidates_have_visible_text(candidates: tuple[dict[str, Any], ...] | None)
     return any(isinstance(text := candidate.get("text"), str) and text.strip() for candidate in candidates)
 
 
+def _candidates_hit_length(candidates: tuple[dict[str, Any], ...] | None) -> bool:
+    """Return whether any ``n > 1`` candidate aggregate ran out of budget."""
+    return any(candidate.get("finish_reason") == "length" for candidate in candidates or ())
+
+
 async def aclose_with_error_precedence(
     resource: Any,
     *,
@@ -773,13 +778,15 @@ async def suppress_thinking_blocks(
     Closing the wrapper eagerly closes the upstream adapter iterator so client
     cancellation still reaches SGLang's ``/abort_request`` cleanup.
 
-    A stream that terminates without ever producing meaningful visible output
-    (no non-whitespace text on any choice or candidate, no tool calls) is
-    stamped with ``error_code="empty_model_output"`` on the terminal chunk —
-    e.g. when private reasoning consumed the entire generation budget. Without
-    the stamp such a request settles as a nominally successful but unusable
+    A stream that stops without ever producing meaningful visible output (no
+    non-whitespace text on any choice or candidate, no tool calls) is stamped
+    with ``error_code="empty_model_output"`` on the terminal chunk. Without the
+    stamp such a request settles as a nominally successful but unusable
     response that clients cannot distinguish from a real answer (#3104/#3136).
-    Cancelled and already-errored terminals are left untouched.
+    A ``length`` terminal keeps its finish reason even when private reasoning
+    consumed the whole budget: ``length`` already reports the truncation, as
+    the OpenAI contract does for reasoning models. Cancelled and
+    already-errored terminals are left untouched.
     """
     states: dict[int, ThinkingBlockStripper] = {}
     emitted_visible_text = False
@@ -856,10 +863,11 @@ async def suppress_thinking_blocks(
                 and not emitted_meaningful_text
                 and not emitted_tool_call
                 and rewritten.error_code is None
-                and rewritten.finish_reason in ("stop", "length")
+                and rewritten.finish_reason == "stop"
+                and not _candidates_hit_length(rewritten.candidates)
                 and not _candidates_have_visible_text(rewritten.candidates)
             ):
-                reason = " (private reasoning consumed the generation budget)" if hid_reasoning else ""
+                reason = " (the model produced only private reasoning)" if hid_reasoning else ""
                 rewritten = replace(
                     rewritten,
                     finish_reason="error",
diff --git a/packages/sie_server/tests/adapters/test_generation_base.py b/packages/sie_server/tests/adapters/test_generation_base.py
index a2d755695..204ff1f89 100644
--- a/packages/sie_server/tests/adapters/test_generation_base.py
+++ b/packages/sie_server/tests/adapters/test_generation_base.py
@@ -292,8 +292,8 @@ async def test_suppress_thinking_blocks_hides_truncated_reasoning_and_its_logpro
     assert normalized[0].is_first is False
     assert normalized[0].logprobs is None
     assert normalized[1].done is True
-    assert normalized[1].finish_reason == "error"
-    assert normalized[1].error_code == "empty_model_output"
+    assert normalized[1].finish_reason == "length"
+    assert normalized[1].error_code is None
     assert normalized[1].completion_tokens == 2
     assert normalized[1].text_delta == ""
     assert normalized[1].logprobs is None
@@ -474,9 +474,9 @@ async def test_suppress_thinking_blocks_surfaces_standalone_close_failure() -> N
 
 
 @pytest.mark.asyncio
-async def test_reasoning_consuming_the_whole_budget_stamps_empty_output_error() -> None:
-    """#3104/#3136: budget exhausted inside a think block -> only whitespace is
-    visible and the terminal chunk must carry the typed empty-output error.
+async def test_reasoning_consuming_the_whole_budget_keeps_the_length_finish() -> None:
+    """Budget exhausted inside a think block: ``length`` already reports the
+    truncation, so the terminal keeps it instead of an empty-output error.
     """
     source = _chunks(
         GenerationChunk(text_delta="\n\n", is_first=True),
@@ -487,6 +487,28 @@ async def test_reasoning_consuming_the_whole_budget_stamps_empty_output_error()
 
     normalized = [chunk async for chunk in suppress_thinking_blocks(source)]
 
+    assert "".join(chunk.text_delta for chunk in normalized) == "\n\n"
+    terminal = normalized[-1]
+    assert terminal.done is True
+    assert terminal.finish_reason == "length"
+    assert terminal.error_code is None
+    assert terminal.error_message is None
+    assert terminal.completion_tokens == 32
+
+
+@pytest.mark.asyncio
+async def test_reasoning_only_answer_that_stops_stamps_empty_output_error() -> None:
+    """#3104/#3136: a model that stops after only private reasoning must not
+    settle as a nominally successful empty answer.
+    """
+    source = _chunks(
+        GenerationChunk(text_delta="\n\n", is_first=True),
+        GenerationChunk(text_delta="<think>private reasoning</think>"),
+        GenerationChunk(text_delta="", done=True, finish_reason="stop", completion_tokens=32),
+    )
+
+    normalized = [chunk async for chunk in suppress_thinking_blocks(source)]
+
     assert "".join(chunk.text_delta for chunk in normalized) == "\n\n"
     terminal = normalized[-1]
     assert terminal.done is True
@@ -573,14 +595,14 @@ async def test_candidates_with_visible_text_are_not_stamped() -> None:
 
 @pytest.mark.asyncio
 async def test_candidates_all_reasoning_are_stamped() -> None:
-    """Non-streaming n>1 where every candidate was reasoning-only."""
+    """Non-streaming n>1 where every candidate stopped after reasoning only."""
     terminal = GenerationChunk(
         text_delta="",
         done=True,
-        finish_reason="length",
+        finish_reason="stop",
         candidates=(
-            {"index": 0, "text": "<think>truncated reasoning", "finish_reason": "length"},
-            {"index": 1, "text": "<think>more truncated reasoning", "finish_reason": "length"},
+            {"index": 0, "text": "<think>reasoning</think>", "finish_reason": "stop"},
+            {"index": 1, "text": "<think>more reasoning</think>", "finish_reason": "stop"},
         ),
     )
 
@@ -591,6 +613,49 @@ async def test_candidates_all_reasoning_are_stamped() -> None:
     assert "reasoning" in normalized.error_message
 
 
+@pytest.mark.asyncio
+async def test_truncated_reasoning_candidates_keep_the_length_finish() -> None:
+    """Non-streaming n>1 where every candidate ran out of budget while reasoning."""
+    terminal = GenerationChunk(
+        text_delta="",
+        done=True,
+        finish_reason="length",
+        candidates=(
+            {"index": 0, "text": "<think>truncated reasoning", "finish_reason": "length"},
+            {"index": 1, "text": "<think>more truncated reasoning", "finish_reason": "length"},
+        ),
+    )
+
+    [normalized] = [chunk async for chunk in suppress_thinking_blocks(_chunks(terminal))]
+
+    assert normalized.finish_reason == "length"
+    assert normalized.error_code is None
+    assert [candidate["text"] for candidate in normalized.candidates or ()] == ["", ""]
+
+
+@pytest.mark.asyncio
+async def test_a_truncated_candidate_keeps_a_stop_terminal_unstamped() -> None:
+    """SGLang's n>1 terminal reports ``stop`` at the top level; a candidate that
+    ran out of budget during reasoning keeps its own ``length`` instead of
+    turning the whole answer into an empty-output error.
+    """
+    terminal = GenerationChunk(
+        text_delta="",
+        done=True,
+        finish_reason="stop",
+        candidates=(
+            {"index": 0, "text": "<think>truncated reasoning", "finish_reason": "length"},
+            {"index": 1, "text": "<think>reasoning</think>", "finish_reason": "stop"},
+        ),
+    )
+
+    [normalized] = [chunk async for chunk in suppress_thinking_blocks(_chunks(terminal))]
+
+    assert normalized.finish_reason == "stop"
+    assert normalized.error_code is None
+    assert [candidate["finish_reason"] for candidate in normalized.candidates or ()] == ["length", "stop"]
+
+
 @pytest.mark.asyncio
 async def test_collect_generation_carries_terminal_error_fields() -> None:
     """#3104/#3136: buffered aggregation must not drop the terminal chunk's
@@ -710,7 +775,8 @@ async def test_suppress_thinking_blocks_keeps_truncated_reasoning_when_asked() -
 
     assert "".join(chunk.reasoning_delta for chunk in kept) == "cut off mid"
     assert "".join(chunk.text_delta for chunk in kept) == ""
-    assert kept[-1].error_code == "empty_model_output"
+    assert kept[-1].finish_reason == "length"
+    assert kept[-1].error_code is None
 
 
 @pytest.mark.asyncio
diff --git a/packages/sie_server/tests/adapters/test_sglang_muse_channels.py b/packages/sie_server/tests/adapters/test_sglang_muse_channels.py
index 3c78179b5..7f9db8770 100644
--- a/packages/sie_server/tests/adapters/test_sglang_muse_channels.py
+++ b/packages/sie_server/tests/adapters/test_sglang_muse_channels.py
@@ -314,7 +314,7 @@ def test_worker_pipeline_converts_arguments_by_their_declared_type() -> None:
     assert json.loads(deltas[1].arguments_delta) == {"days": 3}
 
 
-def test_reasoning_that_uses_the_whole_budget_is_an_explicit_error() -> None:
+def test_reasoning_that_uses_the_whole_budget_keeps_the_length_finish() -> None:
     stream = suppress_thinking_blocks(
         translate_muse_channels(_chunks(" to=self<|message|>Still", finish_reason="length"))
     )
@@ -322,7 +322,8 @@ def test_reasoning_that_uses_the_whole_budget_is_an_explicit_error() -> None:
     chunks = asyncio.run(_collect(stream))
 
     assert _text(chunks) == ""
-    assert chunks[-1].error_code == "empty_model_output"
+    assert chunks[-1].finish_reason == "length"
+    assert chunks[-1].error_code is None
 
 
 def test_adapter_translates_what_the_base_adapter_streams() -> None:
diff --git a/packages/sie_server/tests/processors/test_streaming.py b/packages/sie_server/tests/processors/test_streaming.py
index d65acda55..dd2f7bb76 100644
--- a/packages/sie_server/tests/processors/test_streaming.py
+++ b/packages/sie_server/tests/processors/test_streaming.py
@@ -5008,6 +5008,27 @@ async def test_queued_chat_include_reasoning_publishes_text_and_reasoning_in_mod
 @pytest.mark.asyncio
 async def test_queued_chat_include_reasoning_still_fails_an_answer_that_is_only_reasoning(
     monkeypatch: pytest.MonkeyPatch,
+) -> None:
+    nc, proc = _reasoning_processor(
+        monkeypatch,
+        [
+            GenerationChunk(text_delta="thinking and then stopping", is_first=True),
+            GenerationChunk(text_delta="", done=True, finish_reason="stop", prompt_tokens=1, completion_tokens=8),
+        ],
+    )
+
+    await proc.process(_make_msg(_reasoning_work(stream=True)), "Qwen/Qwen3.5-4B")
+
+    decoded = _decode_chunks(nc)
+    assert "".join(chunk.get("reasoning_delta", "") for chunk in decoded) == "thinking and then stopping"
+    assert not any(chunk.get("text_delta") for chunk in decoded)
+    assert decoded[-1]["done"] is True
+    assert decoded[-1]["error"]["code"] == "empty_model_output"
+
+
+@pytest.mark.asyncio
+async def test_queued_chat_include_reasoning_reports_truncated_reasoning_as_length(
+    monkeypatch: pytest.MonkeyPatch,
 ) -> None:
     nc, proc = _reasoning_processor(
         monkeypatch,
@@ -5023,4 +5044,5 @@ async def test_queued_chat_include_reasoning_still_fails_an_answer_that_is_only_
     assert "".join(chunk.get("reasoning_delta", "") for chunk in decoded) == "thinking until the budget ends"
     assert not any(chunk.get("text_delta") for chunk in decoded)
     assert decoded[-1]["done"] is True
-    assert decoded[-1]["error"]["code"] == "empty_model_output"
+    assert decoded[-1]["finish_reason"] == "length"
+    assert "error" not in decoded[-1]
diff --git a/packages/sie_ts_sdk/README.md b/packages/sie_ts_sdk/README.md
index 346ab773b..a85d1b12f 100644
--- a/packages/sie_ts_sdk/README.md
+++ b/packages/sie_ts_sdk/README.md
@@ -160,8 +160,9 @@ Granite Guardian's shipped risk dimension is `harm`; requesting another
 dimension in prompt text does not change it. Valid thresholded verdicts are
 `Yes` (unsafe) and `No` (safe); missing or invalid verdicts fail with
 `invalid_guard_verdict`. Never interpret an error or empty output as safe.
-Reasoning remains private; a budget exhausted without usable output fails with
-`empty_model_output`.
+Reasoning remains private. A budget exhausted during reasoning returns an empty
+answer whose finish reason is `"length"`; a model that stops without usable output
+fails with `empty_model_output`.
 
 ```typescript
 // Aggregated result