diff --git a/crates/ml/src/trainers/dqn/trainer/mod.rs b/crates/ml/src/trainers/dqn/trainer/mod.rs
index 8376d39f0..5663e8fd7 100644
--- a/crates/ml/src/trainers/dqn/trainer/mod.rs
+++ b/crates/ml/src/trainers/dqn/trainer/mod.rs
@@ -110,6 +110,8 @@ pub(crate) struct BacktrackingState {
     pub lr_boost_remaining: usize,
     /// Original LR (saved before LR boost).
     pub original_lr: f64,
+    /// Previous epoch's val_Sharpe for freeze detection (NOT training Sharpe).
+    pub prev_val_sharpe: f32,
 }
 
 impl BacktrackingState {
@@ -132,6 +134,7 @@ impl BacktrackingState {
             min_improvement_rate: 0.1,  // min 0.1 val_Sharpe/epoch improvement
             best_sharpe: f32::NEG_INFINITY,
             best_epoch: 0,
+            prev_val_sharpe: 0.0,
             temp_boost_remaining: 0,
             lr_boost_remaining: 0,
             original_lr: 0.0,
diff --git a/crates/ml/src/trainers/dqn/trainer/training_loop.rs b/crates/ml/src/trainers/dqn/trainer/training_loop.rs
index 7fd9b92ad..3d745b8e6 100644
--- a/crates/ml/src/trainers/dqn/trainer/training_loop.rs
+++ b/crates/ml/src/trainers/dqn/trainer/training_loop.rs
@@ -575,7 +575,10 @@ impl DQNTrainer {
             ).await?;
 
             // Task 10: Trajectory backtracking -- detect plateau, rewind, perturb
-            if self.run_backtracking_epoch_end(epoch, log_output.epoch_sharpe)? {
+            // Use val_Sharpe (deterministic backtest), NOT epoch_sharpe (training Sharpe).
+            // Training Sharpe oscillates even when the model is frozen, preventing detection.
+            let val_sharpe_for_backtrack = -log_output.val_loss;  // val_loss = -sharpe
+            if self.run_backtracking_epoch_end(epoch, val_sharpe_for_backtrack)? {
                 // PLATEAU_EXHAUSTED -- save best model and exit
                 info!(
                     "PLATEAU_EXHAUSTED: restoring best model from epoch {} before exit",
@@ -2453,27 +2456,18 @@ impl DQNTrainer {
         }
 
         // 2. Save checkpoint when improving
-        // Use second-to-last sharpe (last is current epoch, already pushed before this runs)
-        let prev_sharpe = if self.sharpe_history.len() >= 2 {
-            self.sharpe_history[self.sharpe_history.len() - 2] as f32
-        } else {
-            0.0
-        };
-        let improvement_rate = val_sharpe_f32 - prev_sharpe; // per epoch
+        // Use prev_val_sharpe from backtracking state (tracks val_Sharpe, not training Sharpe)
+        let improvement_rate = val_sharpe_f32 - self.backtracking.prev_val_sharpe;
         if improvement_rate > 0.01 && !self.backtracking.route_active {
             self.save_backtracking_checkpoint(epoch, improvement_rate, val_sharpe_f32)?;
         }
 
         // 3. Detect plateau: BOTH Q-gap frozen AND val_Sharpe stagnating.
-        // OR alone is too aggressive — stable Q-gap + rising val_Sharpe is healthy convergence,
-        // not a plateau. Only intervene when NOTHING improves.
+        // Uses val_Sharpe (deterministic backtest), NOT training Sharpe.
+        // Training Sharpe oscillates even when the model is truly frozen.
         let q_gap_frozen = self.is_q_gap_frozen();
-        let sharpe_frozen = if self.sharpe_history.len() >= 2 {
-            let prev = self.sharpe_history[self.sharpe_history.len() - 2] as f32;
-            (val_sharpe_f32 - prev).abs() < 0.01
-        } else {
-            false
-        };
+        let sharpe_frozen = (val_sharpe_f32 - self.backtracking.prev_val_sharpe).abs() < 0.01;
+        self.backtracking.prev_val_sharpe = val_sharpe_f32;  // update for next epoch
         if q_gap_frozen && sharpe_frozen {
             self.backtracking.plateau_epochs += 1;
             if self.backtracking.plateau_epochs > 0 {