diff --git a/crates/ml/tests/dqn_early_stopping_termination_test.rs b/crates/ml/tests/dqn_early_stopping_termination_test.rs
index 25f85681c..4b49b0226 100644
--- a/crates/ml/tests/dqn_early_stopping_termination_test.rs
+++ b/crates/ml/tests/dqn_early_stopping_termination_test.rs
@@ -147,13 +147,14 @@ async fn test_early_stopping_terminates_with_error() {
     hyperparams.epochs = 10;
     hyperparams.gradient_collapse_multiplier = 1e9;
     hyperparams.gradient_collapse_patience = 3;
+    hyperparams.min_epochs_before_stopping = 1; // Allow collapse at any epoch
     hyperparams.checkpoint_frequency = 1;
     hyperparams.batch_size = 32;
-    hyperparams.buffer_size = 1024;  // Minimum for GPU PER (MIN_GPU_CAPACITY)
+    hyperparams.buffer_size = 1024;
     hyperparams.min_replay_size = 32;
     hyperparams.warmup_steps = 0;
-    hyperparams.max_training_steps_per_epoch = 300;  // Fast epochs: clears warmup (300>204), ~3s vs ~370s
-    hyperparams.replay_buffer_vram_fraction = 0.0;  // Disable AutoReplaySizer for test determinism
+    hyperparams.max_training_steps_per_epoch = 300;
+    hyperparams.replay_buffer_vram_fraction = 0.0;
 
     hyperparams.gpu_n_episodes = 16;
     hyperparams.gpu_timesteps_per_episode = 50;
@@ -218,12 +219,13 @@ async fn test_gradient_collapse_propagates_error() {
     hyperparams.epochs = 10;
     hyperparams.gradient_collapse_multiplier = 1e9;
     hyperparams.gradient_collapse_patience = 3;
+    hyperparams.min_epochs_before_stopping = 1;
     hyperparams.batch_size = 32;
-    hyperparams.buffer_size = 1024;  // Minimum for GPU PER (MIN_GPU_CAPACITY)
+    hyperparams.buffer_size = 1024;
     hyperparams.min_replay_size = 32;
-    hyperparams.warmup_steps = 0;  // Skip train_step() warmup so training_steps increments immediately
-    hyperparams.max_training_steps_per_epoch = 300;  // Fast epochs: clears warmup (300>204), ~3s vs ~370s
-    hyperparams.replay_buffer_vram_fraction = 0.0;  // Disable AutoReplaySizer for test determinism
+    hyperparams.warmup_steps = 0;
+    hyperparams.max_training_steps_per_epoch = 300;
+    hyperparams.replay_buffer_vram_fraction = 0.0;
 
     hyperparams.gpu_n_episodes = 16;
     hyperparams.gpu_timesteps_per_episode = 50;
diff --git a/crates/ml/tests/smoke_test_real_data.rs b/crates/ml/tests/smoke_test_real_data.rs
index 25b6281b8..b56eb4615 100644
--- a/crates/ml/tests/smoke_test_real_data.rs
+++ b/crates/ml/tests/smoke_test_real_data.rs
@@ -707,7 +707,7 @@ mod gpu_smoke {
             timesteps_per_episode: timesteps,
             total_bars: num_bars as i32,
             episode_length: timesteps,
-            epsilon: 0.0, // NoisyNets handle exploration
+            epsilon: 0.1, // 10% random ensures ≥2 distinct actions in 200 samples
             gamma: 0.99,
             // sigma must be large enough that NoisyNet noise dominates Q-value gaps
             // even on wide networks (256-dim). With random Xavier init on a 256-wide