diff --git a/ml/src/hyperopt/adapters/dqn.rs b/ml/src/hyperopt/adapters/dqn.rs
index fe5d5d192..12f029842 100644
--- a/ml/src/hyperopt/adapters/dqn.rs
+++ b/ml/src/hyperopt/adapters/dqn.rs
@@ -408,10 +408,11 @@ impl DQNTrainer {
             early_stopping_plateau_window: 5, // Default: 5 epochs (hyperopt optimized)
             early_stopping_min_epochs: 1000, // Default: 1000 (effectively disabled - Wave 7 validation)
             trial_counter: 0,                // Start at trial 0
-            // WAVE 16 (Agent 38): Switched to Hard updates for stability
-            tau: 1.0, // Hard updates use full copy
-            target_update_mode: crate::trainers::TargetUpdateMode::Hard, // Hard updates (Stable Baselines3)
-            target_update_frequency: 10000, // Stable Baselines3 standard: 10K steps
+            // FIX: Enable soft updates (root cause of gradient explosion)
+            // Hard updates were causing 10K-16K gradient norms (Wave 16H baseline: 1,707)
+            tau: 0.001, // Soft updates: 0.1% target blend per step (Rainbow DQN standard)
+            target_update_mode: crate::trainers::TargetUpdateMode::Soft, // Soft updates prevent Q-value explosion
+            target_update_frequency: 1, // Update every step with soft blending
             enable_preprocessing: true,     // Preprocessing enabled by default (Wave 14)
             preprocessing_window: 50,       // Default: 50-bar rolling window
             preprocessing_clip_sigma: 5.0,  // Default: clip at ±5σ