diff --git a/config/training/dqn-smoketest.toml b/config/training/dqn-smoketest.toml
index c03875057..ec7924ab0 100644
--- a/config/training/dqn-smoketest.toml
+++ b/config/training/dqn-smoketest.toml
@@ -37,7 +37,7 @@ count_bonus_coefficient = 0.1
 q_gap_threshold = 0.05
 
 [replay_buffer]
-buffer_size = 1024
+buffer_size = 256
 min_replay_size = 64
 regime_replay_decay = 0.3
 
diff --git a/crates/ml-dqn/src/replay_buffer_type.rs b/crates/ml-dqn/src/replay_buffer_type.rs
index 66304bd9d..a690001f1 100644
--- a/crates/ml-dqn/src/replay_buffer_type.rs
+++ b/crates/ml-dqn/src/replay_buffer_type.rs
@@ -176,10 +176,11 @@ impl ReplayBufferType {
         batch_size: usize,
         stream: &Arc<CudaStream>,
     ) -> Result<Self, MLError> {
-        const MIN_GPU_CAPACITY: usize = 1024;
+        // Minimum capacity = batch_size (can't sample more than buffer holds)
+        let min_capacity = batch_size.max(1);
         let mut try_cap = capacity;
 
-        while try_cap >= MIN_GPU_CAPACITY {
+        while try_cap >= min_capacity {
             match Self::new_gpu_prioritized(try_cap, state_dim, alpha, beta, beta_max, beta_annealing_steps, max_memory_bytes, batch_size, stream) {
                 Ok(buf) => {
                     if try_cap < capacity {
@@ -201,14 +202,17 @@ impl ReplayBufferType {
                         try_cap, e, try_cap / 2
                     );
                     try_cap /= 2;
-                    if try_cap < MIN_GPU_CAPACITY {
+                    if try_cap < min_capacity {
                         return Err(e);
                     }
                 }
             }
         }
 
-        Err(MLError::ModelError("GPU PER capacity below floor".into()))
+        Err(MLError::ModelError(format!(
+            "GPU PER allocation failed: capacity {} below minimum {} (batch_size)",
+            try_cap, min_capacity
+        )))
     }
 
     /// Sample a batch from the buffer