diff --git a/sync_rtl/top/mlkem_top.v b/sync_rtl/top/mlkem_top.v
index fe81a4f..f17a1ed 100644
--- a/sync_rtl/top/mlkem_top.v
+++ b/sync_rtl/top/mlkem_top.v
@@ -506,10 +506,20 @@ module mlkem_top #(
     // accumulator source: e_hat[i] for first term (j==0), else running t_hat[i].
     // e_hat[i] lives in bank_se at relative slot (slot_e_rt-slot_s_rt + m_i) = K+m_i.
     // t_hat[i] lives in bank_t at relative slot m_i.
-    wire [13:0] m_eacc_full = ({2'b0, k_r} + {2'b0, m_i})*256 + m_oidx;  // K+m_i
-    wire [13:0] m_tacc_full = m_i*256 + m_oidx;
-    wire [11:0]  m_acc_src = (m_j == 2'd0) ? bank_se[m_eacc_full[PSE_AW-1:0]]
-                                           : bank_t [m_tacc_full[PT_AW-1:0]];
+    // Registered read-ahead: present the index that the NEXT pm_vo will consume
+    //   m_acc_radr = pm_vo ? m_oidx+1 : m_oidx
+    // and register the selected bank into m_acc_rd, consumed one cycle later when
+    // pm_vo is high. The same-index read-modify-write on bank_t is read-old:
+    // the running t_hat source was written by the previous (j-1) term, long
+    // settled; the value consumed this cycle was registered before this cycle's
+    // write to m_oidx. Cadence is CALC(vo=0)/C0(vo=1)/C1(vo=1).
+    wire [7:0]  m_acc_radr = pm_vo ? (m_oidx + 8'd1) : m_oidx;
+    wire [13:0] m_eacc_full = ({2'b0, k_r} + {2'b0, m_i})*256 + m_acc_radr;  // K+m_i
+    wire [13:0] m_tacc_full = m_i*256 + m_acc_radr;
+    wire [11:0] m_acc_rd_next = (m_j == 2'd0) ? bank_se[m_eacc_full[PSE_AW-1:0]]
+                                              : bank_t [m_tacc_full[PT_AW-1:0]];
+    reg  [11:0] m_acc_rd;
+    wire [11:0]  m_acc_src = m_acc_rd;
     // (a + b) mod Q  (both < Q, sum < 2Q): one conditional subtract
     wire [12:0]  m_sum  = {1'b0, m_acc_src} + {1'b0, pm_coeff};
     wire [11:0]  m_accq = (m_sum >= 13'(Q)) ? (m_sum - 13'(Q)) : m_sum[11:0];
@@ -566,6 +576,7 @@ module mlkem_top #(
             pm_valid   <= 1'b0;
             pm_a_rd    <= 12'd0;
             pm_b_rd    <= 12'd0;
+            m_acc_rd   <= 12'd0;
             e_poly     <= 3'd0;
             e_pair     <= 8'd0;
             e_byte     <= 2'd0;
@@ -752,6 +763,11 @@ module mlkem_top #(
 
             // ---- ST_M: t_hat[i] = e_hat[i] + sum_j A[i][j] o s_hat[j] ----
             if (st == ST_M) begin
+                // registered accumulator read (sd_bram timing): captures the
+                // index the next pm_vo will consume. Free-running within ST_M;
+                // primed during LOAD (m_oidx held at 0).
+                m_acc_rd <= m_acc_rd_next;
+
                 // LOAD: present read-ahead addr to bank_a/bank_se; the pair
                 // registered last cycle (pm_a_rd/pm_b_rd) is consumed by
                 // poly_mul this cycle (pm_valid). poly_mul holds ready high