diff --git a/crates/ml/src/cuda_pipeline/c51_grad_kernel.cu b/crates/ml/src/cuda_pipeline/c51_grad_kernel.cu
index 6a3f4dacd..15a5b03fa 100644
--- a/crates/ml/src/cuda_pipeline/c51_grad_kernel.cu
+++ b/crates/ml/src/cuda_pipeline/c51_grad_kernel.cu
@@ -26,8 +26,7 @@ extern "C" __global__ void c51_grad_kernel(
     const float* __restrict__ branch_scales,    /* [B, 4] per-sample per-branch gradient scale */
     const float* __restrict__ per_sample_support, /* [B, 3] per-sample [v_min, v_max, delta_z] */
     const float* __restrict__ liquid_mod,        /* [4] pinned device-mapped per-branch modulators */
-    const float* __restrict__ atom_positions,    /* [4, num_atoms] adaptive positions. NULL = linear. */
-    const float* __restrict__ q_mean_ema_ptr)   /* [1] pinned device-mapped Q-mean EMA */
+    const float* __restrict__ atom_positions)    /* [4, num_atoms] adaptive positions. NULL = linear. */
 {
     int tid = blockIdx.x * blockDim.x + threadIdx.x;
     int total_elems = batch_size * num_atoms;
@@ -133,11 +132,6 @@ extern "C" __global__ void c51_grad_kernel(
         branch_base += batch_size * A_d * num_atoms;
     }
 
-    /* Q-mean drift: handled adaptively by E1 enrichment (Q-value reality check).
-     * Hardcoded penalty removed — E1 computes bias correction from actual eval
-     * performance and adapts to changing reward scales automatically. */
-    (void)q_mean_ema_ptr; /* parameter kept for ABI compatibility */
-
     /* Single deterministic write — no atomicAdd */
     d_value_logits[b * num_atoms + j] = d_val_sum;
 }
diff --git a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
index ca52ca5a7..498160de3 100644
--- a/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
+++ b/crates/ml/src/cuda_pipeline/gpu_dqn_trainer.rs
@@ -1226,13 +1226,10 @@ pub struct GpuDqnTrainer {
     // No gradient backward — the 4 edges encode structural dependencies (dir→mag, etc).
     // Training these would require a graph-level loss, deferred to future work.
     /// Edge parameters: 4 edges × 15 params (W_gate[6] + W_msg[9]) = 60.
+    /// Fixed at near-identity initialization — not trained. The 4 edges encode
+    /// structural dependencies (dir→mag, etc.) via domain knowledge; training
+    /// would require a graph-level loss (deferred to future work).
     graph_params: CudaSlice<f32>,
-    /// Adam first moment for graph_params [60]. Reserved for future graph-level training.
-    graph_adam_m: CudaSlice<f32>,
-    /// Adam second moment for graph_params [60]. Reserved for future graph-level training.
-    graph_adam_v: CudaSlice<f32>,
-    /// Adam step counter for graph message pass optimizer. Reserved for future graph-level training.
-    graph_adam_step: i32,
     /// branch_graph_message_pass kernel handle.
     graph_msg_kernel: CudaFunction,
 
@@ -5190,10 +5187,6 @@ impl GpuDqnTrainer {
             .map_err(|e| MLError::ModelError(format!("cpbi cubin (graph_msg): {e}")))?;
         let graph_msg_kernel = cpbi_module_graph.load_function("branch_graph_message_pass")
             .map_err(|e| MLError::ModelError(format!("branch_graph_message_pass load: {e}")))?;
-        let graph_adam_m = stream.alloc_zeros::<f32>(60)
-            .map_err(|e| MLError::ModelError(format!("alloc graph_adam_m: {e}")))?;
-        let graph_adam_v = stream.alloc_zeros::<f32>(60)
-            .map_err(|e| MLError::ModelError(format!("alloc graph_adam_v: {e}")))?;
         // Initialize graph_params: W_gate biases near zero, W_msg near identity (0.1 on diagonal).
         // Layout per edge (15 floats): W_gate[6] then W_msg[9 = 3×3 row-major].
         let mut graph_params_host = [0.0_f32; 60];
@@ -5810,9 +5803,6 @@ impl GpuDqnTrainer {
             q_mean_ema_pinned,
             q_mean_ema_dev_ptr,
             graph_params,
-            graph_adam_m,
-            graph_adam_v,
-            graph_adam_step: 0,
             graph_msg_kernel,
             denoise_params,
             denoise_adam_m,
@@ -8666,7 +8656,6 @@ impl GpuDqnTrainer {
                 .arg(&self.per_sample_support_ptr)
                 .arg(&self.liquid_mod_buf.raw_ptr())
                 .arg(&self.atom_positions_buf)
-                .arg(&self.q_mean_ema_dev_ptr)
                 .launch(LaunchConfig {
                     grid_dim: (blocks, 1, 1),
                     block_dim: (256, 1, 1),