diff --git a/src/dynamics/transformed_hamiltonian.rs b/src/dynamics/transformed_hamiltonian.rs
index 77fae05..f560a20 100644
--- a/src/dynamics/transformed_hamiltonian.rs
+++ b/src/dynamics/transformed_hamiltonian.rs
@@ -192,7 +192,7 @@ impl<M: Math> TransformedPoint<M> {
         if !math.array_all_finite(&self.transformed_position) {
             return false;
         }
-        if !math.array_all_finite_and_nonzero(&self.transformed_gradient) {
+        if !math.array_all_finite(&self.transformed_gradient) {
             return false;
         }
         if !math.array_all_finite(&self.untransformed_gradient) {
diff --git a/tests/zero_gradient_init.rs b/tests/zero_gradient_init.rs
new file mode 100644
index 0000000..c7eb98b
--- /dev/null
+++ b/tests/zero_gradient_init.rs
@@ -0,0 +1,91 @@
+/// Regression test: init_state must accept valid initial points whose gradient is exactly zero.
+///
+/// Scenario (mirrors the eight_schools WASM failure):
+///   - Model: standard normal N(0, 1),  dim = 10
+///   - Initial position: all zeros
+///   - Gradient at x = 0: d/dx(-x²/2) = -x = 0  (every component)
+///
+/// Before the fix, `check_all` called `array_all_finite_and_nonzero` on the
+/// transformed gradient, which rejected this mathematically valid point with
+/// `NutsError::BadInitGrad`.  After the fix it only calls `array_all_finite`.
+use std::collections::HashMap;
+
+use nuts_rs::{Chain, CpuLogpFunc, CpuMath, DiagGradNutsSettings, LogpError, Settings};
+use nuts_storable::HasDims;
+use thiserror::Error;
+
+// ---------------------------------------------------------------------------
+// Minimal standard-normal model  (mu = 0, sigma = 1 in every dimension)
+// ---------------------------------------------------------------------------
+
+struct StandardNormal {
+    dim: usize,
+}
+
+#[derive(Debug, Error)]
+enum StandardNormalError {}
+
+impl LogpError for StandardNormalError {
+    fn is_recoverable(&self) -> bool {
+        true
+    }
+}
+
+impl HasDims for StandardNormal {
+    fn dim_sizes(&self) -> HashMap<String, u64> {
+        [
+            ("unconstrained_parameter".to_string(), self.dim as u64),
+        ]
+        .into_iter()
+        .collect()
+    }
+}
+
+impl CpuLogpFunc for StandardNormal {
+    type LogpError = StandardNormalError;
+    type FlowParameters = ();
+    type ExpandedVector = Vec<f64>;
+
+    fn dim(&self) -> usize {
+        self.dim
+    }
+
+    fn logp(&mut self, position: &[f64], grad: &mut [f64]) -> Result<f64, Self::LogpError> {
+        let mut logp = 0f64;
+        for (x, g) in position.iter().copied().zip(grad.iter_mut()) {
+            logp -= x * x / 2.0;
+            *g = -x; // gradient is 0 when x == 0
+        }
+        Ok(logp)
+    }
+
+    fn expand_vector<R: rand::Rng + ?Sized>(
+        &mut self,
+        _rng: &mut R,
+        position: &[f64],
+    ) -> Result<Self::ExpandedVector, nuts_rs::CpuMathError> {
+        Ok(position.to_vec())
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Regression test
+// ---------------------------------------------------------------------------
+
+#[test]
+fn set_position_at_zero_gradient_should_succeed() {
+    let dim = 10;
+    let math = CpuMath::new(StandardNormal { dim });
+
+    let settings = DiagGradNutsSettings::default();
+    let mut rng = rand::rng();
+    let mut chain = settings.new_chain(0, math, &mut rng);
+
+    // All-zero initial position → gradient is identically zero at every component.
+    // This is the mathematically valid mode of N(0,1) in each dimension and must
+    // not be rejected by init_state.
+    let init = vec![0f64; dim];
+    chain
+        .set_position(&init)
+        .expect("set_position should accept a valid point whose gradient is zero");
+}