diff --git a/changelog.d/fix-device-init.fixed.md b/changelog.d/fix-device-init.fixed.md
new file mode 100644
index 0000000..19ad05c
--- /dev/null
+++ b/changelog.d/fix-device-init.fixed.md
@@ -0,0 +1 @@
+Collapse duplicate device-init block in Calibration.__init__. The second copy re-ran the fallback chain, contained an unreachable else branch, and checked `self.device == "cuda"` (a torch.device vs string comparison that never triggered CUDA seeding on the default code path). torch is now seeded uniformly on every path and `torch.cuda.manual_seed_all` is invoked whenever `self.device.type == "cuda"`.
diff --git a/src/microcalibrate/calibration.py b/src/microcalibrate/calibration.py
index 3a2a9ef..28d2bbe 100644
--- a/src/microcalibrate/calibration.py
+++ b/src/microcalibrate/calibration.py
@@ -65,6 +65,9 @@ def __init__(
             sparse_learning_rate (float): Learning rate for the regularizing optimizer. Defaults to 0.2.
             regularize_with_l0 (Optional[bool]): Whether to apply L0 regularization. Defaults to False.
         """
+        # Resolve the torch device exactly once. The fallback chain
+        # (cuda -> mps -> cpu) runs when ``device`` is None so callers
+        # automatically get the best available accelerator.
         if device is not None:
             self.device = torch.device(device)
         else:
@@ -97,17 +100,14 @@ def __init__(
         self.regularize_with_l0 = regularize_with_l0
         self.seed = seed
 
-        if device is not None:
-            self.device = torch.device(device)
+        # Seed torch on every path, and CUDA as well when we actually
+        # resolved to a CUDA device, so stochastic CUDA kernels are
+        # reproducible. ``self.device`` is a torch.device, so comparing
+        # its ``.type`` to the string "cuda" is the correct check.
+        if self.seed is not None:
             torch.manual_seed(self.seed)
-        else:
-            self.device = torch.device(
-                "cuda"
-                if torch.cuda.is_available()
-                else "mps" if torch.mps.is_available() else "cpu"
-            )
-            if self.device == "cuda":
-                torch.cuda.manual_seed(self.seed)
+            if self.device.type == "cuda":
+                torch.cuda.manual_seed_all(self.seed)
 
         self.estimate_matrix = None
         self.targets = None
diff --git a/tests/test_device_init.py b/tests/test_device_init.py
new file mode 100644
index 0000000..724a454
--- /dev/null
+++ b/tests/test_device_init.py
@@ -0,0 +1,82 @@
+"""Regression tests for the duplicated device-init block (finding #6).
+
+Two problems in the old code:
+
+1. Lines 100-110 re-ran the device resolution already done at lines
+   68-75, and the duplicate block's "else" branch was unreachable
+   because the default was "cpu". Users who passed ``device="cuda"``
+   therefore got ``torch.manual_seed`` but never ``torch.cuda.manual_seed``.
+
+2. ``if self.device == "cuda"`` compared a ``torch.device`` to a
+   string, which is technically true in modern torch for CUDA devices
+   but fragile and confusing, and it was inside the unreachable
+   branch anyway.
+"""
+
+import numpy as np
+import pandas as pd
+import torch
+
+from microcalibrate.calibration import Calibration
+
+
+def _make_calibrator(**kwargs) -> Calibration:
+    rng = np.random.default_rng(0)
+    data = pd.DataFrame(
+        {
+            "x": rng.normal(size=20),
+            "y": rng.normal(size=20),
+        }
+    )
+    estimate_matrix = pd.DataFrame(
+        {
+            "sum_x": data["x"],
+            "sum_y": data["y"],
+        }
+    )
+    weights = np.ones(len(data))
+    targets = np.array([data["x"].sum(), data["y"].sum()])
+    return Calibration(
+        estimate_matrix=estimate_matrix,
+        weights=weights,
+        targets=targets,
+        noise_level=0.1,
+        epochs=5,
+        learning_rate=0.01,
+        dropout_rate=0,
+        **kwargs,
+    )
+
+
+def test_device_is_torch_device_not_string() -> None:
+    """``self.device`` must be a torch.device object after init."""
+    calibrator = _make_calibrator(device="cpu")
+    assert isinstance(calibrator.device, torch.device)
+    assert calibrator.device.type == "cpu"
+
+
+def test_cpu_path_seeds_torch() -> None:
+    """Passing ``device='cpu'`` with a seed must still seed torch.
+
+    The previous block lived inside ``if device is not None`` and only
+    called ``torch.manual_seed`` there, but it had a redundant second
+    device resolution. The simplified code seeds torch uniformly on
+    every path.
+    """
+    calibrator = _make_calibrator(device="cpu", seed=123)
+    # Generate a random number after construction; redo the construction
+    # and verify we get the same draw (torch was seeded).
+    draw_a = torch.rand(3)
+
+    calibrator_2 = _make_calibrator(device="cpu", seed=123)
+    draw_b = torch.rand(3)
+
+    assert torch.equal(draw_a, draw_b)
+
+
+def test_default_device_resolves_without_crashing() -> None:
+    """With device=None the fallback chain runs exactly once."""
+    calibrator = _make_calibrator(device=None)
+    # Should resolve to some valid torch.device.
+    assert isinstance(calibrator.device, torch.device)
+    assert calibrator.device.type in {"cuda", "mps", "cpu"}