diff --git a/crates/common/src/observability/logger.rs b/crates/common/src/observability/logger.rs
index 6a98c21b3..48d1798db 100644
--- a/crates/common/src/observability/logger.rs
+++ b/crates/common/src/observability/logger.rs
@@ -190,7 +190,7 @@ pub fn init_json_logger(config: JsonLoggerConfig) -> CommonResult<()> {
             .with_target(true)
             .with_current_span(true)
             .with_span_list(false)
-            .with_writer(std::io::stdout))
+            .with_writer(std::io::stderr))
     } else {
         None
     };
diff --git a/crates/ml/src/trainers/dqn/trainer/training_loop.rs b/crates/ml/src/trainers/dqn/trainer/training_loop.rs
index 4b1fa6f11..d086bde3f 100644
--- a/crates/ml/src/trainers/dqn/trainer/training_loop.rs
+++ b/crates/ml/src/trainers/dqn/trainer/training_loop.rs
@@ -903,7 +903,7 @@ impl DQNTrainer {
                     state_dim,
                     self.hyperparams.gpu_timesteps_per_episode,
                 );
-                optimal.max(32).min(256)
+                optimal.max(32).min(16384)
             }
             Err(_) => 256, // safe fallback
         }
@@ -971,7 +971,7 @@ impl DQNTrainer {
                         aligned_sd,
                         self.hyperparams.gpu_timesteps_per_episode,
                     );
-                    let chosen = optimal.max(32).min(256) as i32; // TEMP: cap at 256 to debug H100 hang
+                    let chosen = optimal.max(32).min(16384) as i32;
                     info!(
                         "GPU auto-scaled n_episodes: {} (SMs={}, VRAM={:.0}MB)",
                         chosen, hw.sm_count, hw.free_memory_mb