diff --git a/README.md b/README.md
index 4200c5f..2c0b043 100644
--- a/README.md
+++ b/README.md
@@ -94,6 +94,13 @@ Workers return a `WorkerResponse`:
 Any other field a worker returns is dropped. `internal` is the one exception, so
 a worker cannot reach a caller with a field nobody agreed to.
 
+A worker raises `ValidationError` for a bad request (400) and
+`FatalWorkerError` when the process itself can no longer serve, such as after a
+CUDA error that poisons the context. A fatal error still fails its own request
+with 500, then `/health` and `/inference` answer 503 and the server exits, so
+the orchestrator restarts the container instead of routing more jobs to it.
+Raise it `from` the original error: the response carries the whole chain.
+
 The `/health` endpoint reports the same `worker` object, supplied through
 `WORKER_NAME` and `WORKER_VERSION`, and available GPU metadata in addition to
 `ok`. Deployments should set `WORKER_VERSION` to the exact worker image tag.
diff --git a/maestro_worker_python/response.py b/maestro_worker_python/response.py
index e860596..3b77c00 100644
--- a/maestro_worker_python/response.py
+++ b/maestro_worker_python/response.py
@@ -23,3 +23,7 @@ class WorkerResponse(BaseModel):
 class ValidationError(Exception):
     def __init__(self, reason):
         self.reason = reason
+
+
+class FatalWorkerError(Exception):
+    """This process can no longer serve; raise it from the error that broke it."""
diff --git a/maestro_worker_python/serve.py b/maestro_worker_python/serve.py
index 5a1964b..3a82aae 100644
--- a/maestro_worker_python/serve.py
+++ b/maestro_worker_python/serve.py
@@ -22,7 +22,7 @@
 from .kill_process import kill_child_processes, terminate_current_process
 from .load_worker import load_worker
 from .request_logging import register_client_safe_request_extractor
-from .response import ValidationError, WorkerResponse
+from .response import FatalWorkerError, ValidationError, WorkerResponse
 
 
 def filter_transactions(event, hint):
@@ -98,14 +98,19 @@ async def lifespan(_app: FastAPI) -> AsyncIterator[None]:
 
 error_counter = 0
 lock = asyncio.Lock()
+# Set once a worker raises FatalWorkerError; there is no way back to healthy.
+fatal_error: str | None = None
+
+
+def _error_body(exc: Exception) -> dict:
+    return {"error": "".join(traceback.format_exception(None, exc, exc.__traceback__))}
 
 
 @app.exception_handler(500)
 async def internal_exception_handler(request: Request, exc: Exception):
     global error_counter
-    tb = "".join(traceback.format_exception(None, exc, exc.__traceback__))
     try:
-        return JSONResponse(status_code=500, content=jsonable_encoder({"error": tb}))
+        return JSONResponse(status_code=500, content=jsonable_encoder(_error_body(exc)))
     finally:
         if error_counter > 10:
             logging.error("Too many consecutive errors, shutting down worker")
@@ -115,6 +120,22 @@ async def internal_exception_handler(request: Request, exc: Exception):
             error_counter += 1
 
 
+@app.exception_handler(FatalWorkerError)
+async def fatal_worker_error_handler(request: Request, exc: FatalWorkerError):
+    global fatal_error
+    fatal_error = str(exc)
+    logging.critical("Worker reported a fatal error, shutting down worker: %s", fatal_error)
+    try:
+        return JSONResponse(status_code=500, content=jsonable_encoder(_error_body(exc)))
+    finally:
+        # uvicorn shuts down gracefully, so this response is still delivered.
+        terminate_current_process()
+
+
+def _unavailable() -> JSONResponse:
+    return JSONResponse(status_code=503, content={"ok": False, "error": fatal_error})
+
+
 @app.exception_handler(ValidationError)
 async def validation_error_handler(request: Request, exc: ValidationError):
     return JSONResponse(status_code=400, content=jsonable_encoder({"error": exc.reason}))
@@ -153,6 +174,8 @@ def _with_worker_identity(result):
 @app.post("/inference", response_model=WorkerResponse)
 async def inference(request: Request):
     global error_counter
+    if fatal_error is not None:
+        return _unavailable()
     params = await request.json()
     result = await run_in_threadpool(model.inference, input_data=params)
     async with lock:
@@ -167,4 +190,6 @@ async def index(request: Request):
 
 @app.get("/health")
 async def health(request: Request):
+    if fatal_error is not None:
+        return _unavailable()
     return {"ok": True, **get_health_metadata()}
diff --git a/pyproject.toml b/pyproject.toml
index c835820..8353644 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -1,6 +1,6 @@
 [project]
 name = "maestro-worker-python"
-version = "5.1.2"
+version = "5.2.0"
 description = "Utility to run workers on Moises/Maestro"
 readme = "README.md"
 requires-python = ">=3.10,<3.14"
diff --git a/tests/test_serve.py b/tests/test_serve.py
index 7e9bf9f..7ff4cd7 100644
--- a/tests/test_serve.py
+++ b/tests/test_serve.py
@@ -284,3 +284,60 @@ def test_inference_and_health_report_the_identity_over_http(tmp_path, monkeypatc
     assert inference["billable_seconds"] == 2.0
     # Both surfaces read one Settings field each, so they cannot disagree.
     assert health["worker"] == inference["worker"]
+
+
+def _serve_failing_worker(tmp_path, monkeypatch, raised: str):
+    worker_path = tmp_path / "worker.py"
+    worker_path.write_text(
+        "from maestro_worker_python.response import FatalWorkerError, ValidationError\n"
+        "class MoisesWorker:\n"
+        "    calls = 0\n"
+        "    def inference(self, input_data):\n"
+        "        MoisesWorker.calls += 1\n"
+        f"        {raised}\n"
+    )
+    serve_module = _import_serve(monkeypatch, worker_path)
+    terminations: list[bool] = []
+    monkeypatch.setattr(serve_module, "terminate_current_process", lambda: terminations.append(True))
+    client = TestClient(serve_module.app, raise_server_exceptions=False)
+    return serve_module, client, terminations
+
+
+def test_a_fatal_worker_error_fails_the_request_then_stops_the_process_serving(tmp_path, monkeypatch):
+    serve_module, client, terminations = _serve_failing_worker(
+        tmp_path,
+        monkeypatch,
+        'raise FatalWorkerError("CUDA context lost") from RuntimeError("CUDA error: an illegal memory access")',
+    )
+
+    failed = client.post("/inference", json={})
+    assert failed.status_code == 500
+    # The cause is what the job owner needs to debug the fault.
+    assert "an illegal memory access" in failed.json()["error"]
+    assert terminations == [True]
+
+    health = client.get("/health")
+    assert health.status_code == 503
+    assert health.json()["ok"] is False
+
+    refused = client.post("/inference", json={})
+    assert refused.status_code == 503
+    assert serve_module.model.calls == 1
+    assert terminations == [True]
+
+
+@pytest.mark.parametrize(
+    ("raised", "status_code"),
+    [
+        pytest.param('raise RuntimeError("shape mismatch")', 500, id="ordinary-failure"),
+        pytest.param('raise ValidationError("input is too big")', 400, id="validation-error"),
+    ],
+)
+def test_a_recoverable_failure_keeps_the_process_serving(tmp_path, monkeypatch, raised, status_code):
+    serve_module, client, terminations = _serve_failing_worker(tmp_path, monkeypatch, raised)
+
+    assert client.post("/inference", json={}).status_code == status_code
+    assert client.get("/health").status_code == 200
+    assert client.post("/inference", json={}).status_code == status_code
+    assert serve_module.model.calls == 2
+    assert terminations == []
diff --git a/uv.lock b/uv.lock
index 6258f4d..96fca76 100644
--- a/uv.lock
+++ b/uv.lock
@@ -220,7 +220,7 @@ wheels = [
 
 [[package]]
 name = "maestro-worker-python"
-version = "5.1.2"
+version = "5.2.0"
 source = { editable = "." }
 dependencies = [
     { name = "fastapi" },