diff --git a/crates/ml-alpha/examples/alpha_rl_train.rs b/crates/ml-alpha/examples/alpha_rl_train.rs
index df1adbc4e..fdac44db2 100644
--- a/crates/ml-alpha/examples/alpha_rl_train.rs
+++ b/crates/ml-alpha/examples/alpha_rl_train.rs
@@ -885,7 +885,7 @@ fn main() -> Result<()> {
                          "stale":    isv[RL_LR_V_STEPS_SINCE_BEST_INDEX],
                          "warmup":   isv[RL_LR_V_WARMUP_COUNTER_INDEX] },
             },
-            "replay_len": 0usize,
+            "replay_len": trainer.gpu_replay.capacity.min(step + 1),
             // Post-scale, POST-clamp reward stats — what V regression
             // and Q distributional projection actually saw this step.
             // The `scaled_pre_clamp_max` field (below) is the SAME
@@ -1152,7 +1152,7 @@ fn main() -> Result<()> {
                 isv[RL_PPO_CLIP_INDEX],
                 isv[RL_PER_ALPHA_INDEX],
                 isv[RL_REWARD_SCALE_INDEX],
-                0usize,
+                trainer.gpu_replay.capacity.min(step + 1),
                 done_count,
                 reward_sum,
                 sps,
@@ -1298,7 +1298,7 @@ fn main() -> Result<()> {
         summary.final_l_aux = s.l_aux;
         summary.final_l_total = s.l_total;
     }
-    summary.final_replay_len = 0usize;
+    summary.final_replay_len = trainer.gpu_replay.capacity.min(cli.n_steps);
     summary.completed_clean = summary.n_steps_completed == summary.n_steps_planned;
     write_summary(&cli.out, &summary)?;