diff --git a/config/training/dqn-hyperopt.toml b/config/training/dqn-hyperopt.toml
index af9e33cc9..8d5933a4d 100644
--- a/config/training/dqn-hyperopt.toml
+++ b/config/training/dqn-hyperopt.toml
@@ -39,7 +39,7 @@ tau = [0.005, 0.01]                    # log scale in adapter
 hidden_dim_base = [128, 256]           # capped at production default: 512 is 4x slower for marginal benefit
 
 # CQL regularization
-cql_alpha = [0.0, 1.0]
+cql_alpha = [0.5, 5.0]
 
 # Training dynamics
 lr_decay_type = [0, 2]                 # discrete: 0=constant, 1=linear, 2=cosine
diff --git a/config/training/dqn-localdev.toml b/config/training/dqn-localdev.toml
index 7f253a3df..6110c45c5 100644
--- a/config/training/dqn-localdev.toml
+++ b/config/training/dqn-localdev.toml
@@ -62,7 +62,7 @@ n_steps = 3
 tau = 0.005
 c51_warmup_epochs = 5
 her_ratio = 0.2
-cql_alpha = 0.1
+cql_alpha = 1.0
 curiosity_weight = 0.1
 iqn_lambda = 0.25
 spectral_norm_sigma_max = 1.5
diff --git a/config/training/dqn-production.toml b/config/training/dqn-production.toml
index 9633c2e36..67ca3e1f6 100644
--- a/config/training/dqn-production.toml
+++ b/config/training/dqn-production.toml
@@ -67,7 +67,7 @@ n_steps = 3
 tau = 0.005
 c51_warmup_epochs = 5
 her_ratio = 0.2
-cql_alpha = 0.1
+cql_alpha = 1.0
 curiosity_weight = 0.1
 iqn_lambda = 0.25
 spectral_norm_sigma_max = 1.5
diff --git a/config/training/dqn-smoketest.toml b/config/training/dqn-smoketest.toml
index 3ff400dc6..2cb69323b 100644
--- a/config/training/dqn-smoketest.toml
+++ b/config/training/dqn-smoketest.toml
@@ -66,7 +66,7 @@ n_steps = 3
 tau = 0.005
 c51_warmup_epochs = 5
 her_ratio = 0.2
-cql_alpha = 0.1
+cql_alpha = 1.0
 curiosity_weight = 0.1
 iqn_lambda = 0.25
 spectral_norm_sigma_max = 1.5
diff --git a/crates/ml/src/trainers/dqn/config.rs b/crates/ml/src/trainers/dqn/config.rs
index 652b5adaa..d938b9503 100644
--- a/crates/ml/src/trainers/dqn/config.rs
+++ b/crates/ml/src/trainers/dqn/config.rs
@@ -1514,7 +1514,7 @@ impl DQNHyperparameters {
             adam_epsilon: 1e-8, // f32 Adam (master weights are f32)
 
             // Conservative Q-Learning (CQL)
-            cql_alpha: 0.1,                       // Default: mild conservatism (0.1 of 0.0-1.0 range)
+            cql_alpha: 1.0,                       // Strong conservatism — penalizes OOS-destructive Q-values
 
             // QR-DQN (complementary to C51 — IQN for quantile estimation)
             num_quantiles: 32,                    // Default: 32 quantiles
diff --git a/crates/ml/src/trainers/dqn/fused_training.rs b/crates/ml/src/trainers/dqn/fused_training.rs
index 9e3c5517c..70dd7600b 100644
--- a/crates/ml/src/trainers/dqn/fused_training.rs
+++ b/crates/ml/src/trainers/dqn/fused_training.rs
@@ -46,9 +46,9 @@ use super::DQNHyperparameters;
 
 /// Per-component gradient norm budget fractions for auxiliary objectives.
 /// C51 gets whatever remains: `1.0 - sum(active_auxiliary_budgets)`.
-/// When all auxiliaries are active: C51=70%, CQL=15%, IQN=10%, Ens=5%.
-/// When auxiliaries are disabled, their budget automatically goes to C51.
-pub(crate) const CQL_GRAD_BUDGET: f32 = 0.15;
+/// When all auxiliaries are active: C51=60%, CQL=25%, IQN=10%, Ens=5%.
+/// CQL at 25% (was 15%) to enforce conservatism on OOS-destructive Q-values.
+pub(crate) const CQL_GRAD_BUDGET: f32 = 0.25;
 pub(crate) const IQN_GRAD_BUDGET: f32 = 0.10;
 pub(crate) const ENS_GRAD_BUDGET: f32 = 0.05;