diff --git a/crates/ml/examples/hyperopt_baseline_rl.rs b/crates/ml/examples/hyperopt_baseline_rl.rs
index e509c736d..afcdd1173 100644
--- a/crates/ml/examples/hyperopt_baseline_rl.rs
+++ b/crates/ml/examples/hyperopt_baseline_rl.rs
@@ -75,9 +75,9 @@
 //! Hyperopt RL Runner for DQN/PPO on Real Databento Market Data
 //!
 //! Runs hyperparameter optimization using PSO/TPE for DQN, PPO, or both models
-//! on downloaded Databento futures data. Supports two-phase optimization:
+//! on downloaded Databento futures data. Supports four-phase optimization:
 //!
-//! ## Two-Phase Workflow (recommended)
+//! ## Four-Phase Workflow (recommended)
 //!
 //! ```bash
 //! # Phase 1 (default): fix small architecture, search learning dynamics (~15D, ~30 min)
@@ -195,10 +195,11 @@ struct Args {
     #[arg(long)]
     trades_data_dir: Option<PathBuf>,
 
-    /// Three-phase hyperopt phase selection:
+    /// Four-phase hyperopt phase selection:
     ///   fast   — Phase 1: fix architecture + reward, search dynamics (~17D). DEFAULT.
     ///   full   — Phase 2: fix dynamics, search architecture (~5D).
     ///   reward — Phase 3: fix dynamics + architecture, search reward weights only (7D, fastest).
+    ///   risk   — Phase 4: fix dynamics + architecture, search risk params (8D).
     #[arg(long, default_value = "fast")]
     phase: String,
 
@@ -538,7 +539,8 @@ fn main() -> Result<()> {
             "fast" => HyperoptPhase::Fast,
             "full" => HyperoptPhase::Full(load_best_vec("full")),
             "reward" => HyperoptPhase::Reward(load_best_vec("reward")),
-            other => panic!("Invalid --phase '{}'. Must be 'fast', 'full', or 'reward'.", other),
+            "risk" => HyperoptPhase::Risk(load_best_vec("risk")),
+            other => panic!("Invalid --phase '{}'. Must be 'fast', 'full', 'reward', or 'risk'.", other),
         };
         info!("Hyperopt phase: {:?}", args.phase);
         set_hyperopt_phase(phase);
diff --git a/crates/ml/src/hyperopt/adapters/dqn.rs b/crates/ml/src/hyperopt/adapters/dqn.rs
index b0f31cf8b..71a8d6bb9 100644
--- a/crates/ml/src/hyperopt/adapters/dqn.rs
+++ b/crates/ml/src/hyperopt/adapters/dqn.rs
@@ -615,6 +615,14 @@ impl ParameterSpace for DQNParams {
                 tracing::info!("Phase REWARD: fixing {} dynamics+architecture params, searching 7 reward dims", best_vec.len());
                 // Bounds will be applied after vec construction below.
             }
+            HyperoptPhase::Risk(ref best_vec) => {
+                // Phase 4: fix dynamics + architecture from Phase 2 best params.
+                // Search ONLY risk params (8D):
+                //   kelly_fractional (17), dd_threshold (35), loss_aversion (36),
+                //   time_decay_rate (37), q_gap_threshold (38), w_dsr (31), w_pnl (32), w_dd (33).
+                tracing::info!("Phase RISK: fixing {} dynamics+architecture params, searching 8 risk dims", best_vec.len());
+                // Bounds will be applied after vec construction below.
+            }
         }
 
         let mut bounds = vec![
@@ -683,6 +691,19 @@ impl ParameterSpace for DQNParams {
             }
         }
 
+        // Phase RISK: fix ALL dims EXCEPT risk params (8D).
+        // Risk dims: kelly_fractional (17), w_dsr (31), w_pnl (32), w_dd (33),
+        //            dd_threshold (35), loss_aversion (36), time_decay_rate (37),
+        //            q_gap_threshold (38).
+        if let HyperoptPhase::Risk(ref best_vec) = phase {
+            let risk_dims: &[usize] = &[17, 31, 32, 33, 35, 36, 37, 38];
+            for (i, bv) in best_vec.iter().enumerate() {
+                if i < bounds.len() && !risk_dims.contains(&i) {
+                    bounds[i] = (*bv, *bv);
+                }
+            }
+        }
+
         bounds
     }
 
@@ -2626,7 +2647,7 @@ impl HyperparameterOptimizable for DQNTrainer {
             // penalize VRAM estimate for replay buffer capacity.
             let vram_buffer = if budget.gpu_memory_mb <= 8192 { 0 } else { clamped_buffer_size };
             // Aligned state_dim: 64 with OFI (58→64), 56 without (50→56)
-            let aligned_state_dim: usize = if self.mbp10_data_dir.is_some() { 64 } else { 56 };
+            let aligned_state_dim: usize = if self.mbp10_data_dir.is_some() { 80 } else { 72 };
             let vram_check = budget.trial_fits_vram(
                 params.hidden_dim_base,
                 params.batch_size,
diff --git a/crates/ml/src/training_profile.rs b/crates/ml/src/training_profile.rs
index a843fb472..446aff03d 100644
--- a/crates/ml/src/training_profile.rs
+++ b/crates/ml/src/training_profile.rs
@@ -294,6 +294,10 @@ pub enum HyperoptPhase {
     /// Phase 3: fix dynamics + architecture from Phase 2 JSON, search only reward weights (7D).
     /// Fastest phase (~10s/trial) — only experience collection changes, not network or training.
     Reward(Vec<f64>),
+    /// Phase 4: fix dynamics + architecture from Phase 2, search risk params (8D).
+    /// Risk dims: kelly_fractional, dd_threshold, loss_aversion, time_decay_rate,
+    /// q_gap_threshold, w_dsr, w_pnl, w_dd.
+    Risk(Vec<f64>),
 }
 
 impl Default for HyperoptPhase {