--- a/src/models/qwen35moe_forward_cuda.c
+++ b/src/models/qwen35moe_forward_cuda.c
@@ -5859,6 +5859,10 @@
         float krow[PN_QWEN35MOE_ATTN_HEAD_DIM_MAX];
         float vrow[PN_QWEN35MOE_ATTN_HEAD_DIM_MAX];
         if (!scores) return -1;
+        /* CAUSALITY RED ARM (tests/causality_red): the first pass stores every row of the
+         * batch, the second re-attends with each row reading the WHOLE batch -- rows read
+         * K/V written by later rows. Run with QWEN35_FLASH=0 so this path is taken. */
+        for (int red_pass = 0; red_pass < 2; red_pass++)
         for (int t = 0; t < n; t++) {
             int pos = pos0 + t;
             int cache_pos = pos % max_seq;
@@ -5884,7 +5888,7 @@
                 }
             }
 
-            int seq_len = (pos + 1 < max_seq) ? pos + 1 : max_seq;
+            int seq_len = (pos0 + n < max_seq) ? pos0 + n : max_seq;
             const float *qhost = hb_q + (size_t)t * attn_dim;
             float *aout = aout_all + (size_t)t * attn_dim;
             for (int qh = 0; qh < n_head; qh++) {