diff --git a/crates/ml/src/trainers/dqn/trainer/action.rs b/crates/ml/src/trainers/dqn/trainer/action.rs
index 4db11fcd9..4e7171fbf 100644
--- a/crates/ml/src/trainers/dqn/trainer/action.rs
+++ b/crates/ml/src/trainers/dqn/trainer/action.rs
@@ -164,8 +164,18 @@ impl DQNTrainer {
         use rand::rngs::StdRng;
         let epsilon = self.get_epsilon().await? as f32;
         let mut rng = StdRng::seed_from_u64(0xAC7_DEF0);
-        if rng.gen::<f32>() < epsilon { Ok(rng.gen_range(0..5)) }
-        else {
+        if rng.gen::<f32>() < epsilon {
+            // Task 2.5 Bug #2: sample 4-branch factored action space
+            // (dir × mag × ord × urg = 3×3×3×3 = 81), matching MEMORY.md
+            // DQN architecture (updated 2026-04-08). The legacy 0..5 range
+            // predated the 4-branch refactor and was a latent foot-gun.
+            // Encoding: action = dir*27 + mag*9 + ord*3 + urg.
+            let dir = rng.gen_range(0..3usize);
+            let mag = rng.gen_range(0..3usize);
+            let ord = rng.gen_range(0..3usize);
+            let urg = rng.gen_range(0..3usize);
+            Ok(dir * 27 + mag * 9 + ord * 3 + urg)
+        } else {
             let agent = self.agent.read().await;
             let indices = agent.batch_greedy_actions(state)?;
             let best_action = *indices.first().ok_or_else(|| anyhow::anyhow!("argmax on empty Q-values"))? as usize;