diff --git a/DEV_NOTES.md b/DEV_NOTES.md
new file mode 100644
index 0000000..ad4d7c0
--- /dev/null
+++ b/DEV_NOTES.md
@@ -0,0 +1,39 @@
+# Development Notes
+
+Setup friction encountered on this fork during GPU development — kept
+here separately so the main README stays clean. Not universal microsim
+guidance, just what was hit on this machine.
+
+## GPU backend (cupy) setup
+
+1. **numpy version conflict** — microsim's metadata pins `numpy<2.0`, but
+   `cupy-cuda12x` requires `numpy>=2.0`. In practice, numpy 2.4.6 worked
+   fine at runtime with this version of microsim despite the metadata
+   warning — if `pip` flags the conflict, it's likely safe to ignore,
+   but verify by running a real simulation before trusting it.
+
+2. **CUDA headers not found** (`RuntimeError: Failed to find CUDA headers`)
+   — if your system has multiple CUDA toolkit installs (common after
+   driver updates over time), cupy's JIT compiler may not find headers
+   even with `CUDA_PATH` set correctly. Installing the toolkit extra
+   directly into the venv sidesteps the problem:
+```bash
+   pip install "cupy-cuda12x[ctk]"
+```
+
+3. **zarr / numcodecs mismatch** (`ImportError: cannot import name
+   'cbuffer_sizes' from 'numcodecs.blosc'`) — zarr 2.17.x expects an
+   older numcodecs API:
+```bash
+   pip install "numcodecs<0.13"
+```
+
+4. **COSEM data fetching** — loading real segmentation data (e.g. the
+   `cosem` example) requires the optional extra:
+```bash
+   pip install "microsim[cosem]"
+```
+
+None of these are bugs in microsim itself, except where already filed
+as upstream PRs (#140, #141) — mostly environment/dependency friction
+from multiple existing CUDA/Python installs on a dev machine.
diff --git a/README.md b/README.md
index 8bf10c2..5b68f64 100644
--- a/README.md
+++ b/README.md
@@ -64,6 +64,9 @@ These are not installed by default, see the
 or [cupy](https://docs.cupy.dev/en/stable/install.html) installation instructions,
 paying attention to your GPU requirements.  Support for torch is planned.
 
+See [DEV_NOTES.md](DEV_NOTES.md) for real setup issues hit on this fork
+(numpy/cupy version conflicts, CUDA header errors, zarr/numcodecs mismatch).
+
 ## Usage
 
 Construct and run a
diff --git a/src/microsim/schema/modality/_simple_psf.py b/src/microsim/schema/modality/_simple_psf.py
index 586fcb8..de7c4d5 100644
--- a/src/microsim/schema/modality/_simple_psf.py
+++ b/src/microsim/schema/modality/_simple_psf.py
@@ -208,8 +208,9 @@ def render(
         self,
         truth: xrDataArray,  # (F, Z, Y, X)
         em_rates: xrDataArray,  # (C, F, W)
-        *args: Any,
-        **kwargs: Any,
+        objective_lens: ObjectiveLens,
+        settings: Settings,
+        xp: NumpyAPI,
     ) -> xrDataArray:
         """Render a 3D image of the truth for F fluorophores, in C channels.
 
@@ -217,7 +218,9 @@ def render(
         already convolved with the PSF. Therefore, we simply compute the emission flux
         for each fluorophore and each channel.
         """
-        em_image = em_rates.sum(Axis.W) * truth
+        em_rates = em_rates.sum(Axis.W)
+        em_rates = em_rates.copy(data=xp.asarray(em_rates.data))
+        em_image = em_rates * truth
         return DataArray(
             em_image,
             dims=[Axis.C, Axis.F, Axis.Z, Axis.Y, Axis.X],
diff --git a/tests/test_simulation.py b/tests/test_simulation.py
index bf9f267..2c54616 100644
--- a/tests/test_simulation.py
+++ b/tests/test_simulation.py
@@ -166,3 +166,17 @@ def test_simulation_write_from_ground_truth(tmp_path: Path) -> None:
     )
     sim.run()
     assert (tmp_path / "output.zarr").exists()
+
+
+def test_identity_modality_backend_conversion(np_backend: ms.BackendName) -> None:
+    """Regression test for #124: em_rates must be converted to the active
+    backend before use in Identity.render(), same as truth already is."""
+    sim = ms.Simulation(
+        truth_space=ms.ShapeScaleSpace(shape=(8, 32, 32), scale=(0.2, 0.1, 0.1)),
+        output_space={"downscale": 1},
+        sample=ms.Sample(labels=[MATSLINES]),
+        modality=ms.Identity(),
+        settings=ms.Settings(np_backend=np_backend),
+    )
+    result = sim.optical_image()
+    assert result.shape[1:] == sim.truth_space.shape