diff --git a/single_shot_inference.py b/single_shot_inference.py
index 99ab33b7..762d491e 100644
--- a/single_shot_inference.py
+++ b/single_shot_inference.py
@@ -66,6 +66,7 @@ INVERSE_ROPE = not _args.no_inverse_rope  # If False, skip inverse RoPE on atten
 SKIP_MHC = _args.skip_mhc   # If True, bypass mHC and use simple residual connections (diagnostic)
 MHC_DIAG = False       # If True, print per-layer mHC diagnostics (B_l row/col sums, C_l values)
 GROWTH_DIAG = True     # If True, print per-layer residual growth analysis
+ATTN_DIAG = True      # If True, print per-layer attention entropy
 # When True: applies inverse RoPE at query position → converts absolute→relative
 # When False: leaves relative position encoding intact for output projection
 # DSV4 partial RoPE only affects last 64/512 dims; first 448 are always un-RoPE'd
@@ -507,16 +508,16 @@ def forward_layer(X_l, w, li, cfg, rope_cos, rope_sin,
             attn_weights = torch.softmax(scores_raw.float(), dim=-1).to(torch.bfloat16)
         attn_out = torch.matmul(attn_weights, v_expanded)  # (n_h, T, hd)
         attn_out = attn_out.permute(1, 0, 2)  # (T, n_h, hd)
-        # Diagnostic: check attention entropy (how spread out the attention is)
-        if MHC_DIAG and li < 3:
+        # Diagnostic: check attention entropy and last-position weight
+        if ATTN_DIAG:
             with torch.no_grad():
                 scores = torch.matmul(q_input, k_expanded.transpose(-1, -2)) * scale  # (n_h, T, seq_len)
                 weights = torch.softmax(scores.float(), dim=-1)  # (n_h, 1, seq_len)
-                # For head 0: what positions get the most weight?
                 w0 = weights[0, 0]  # (seq_len,)
                 top3_pos = torch.topk(w0, min(3, seq_len))
                 entropy = -(w0 * (w0 + 1e-10).log()).sum().item()
-                print(f"  L{li} attn: seq_len={seq_len} entropy={entropy:.2f} top3_pos={top3_pos.indices.tolist()} top3_w={top3_pos.values.tolist()}")
+                last_w = w0[-1].item() if seq_len > 0 else 0
+                print(f"  L{li} attn: seq_len={seq_len} entropy={entropy:.2f} last_pos_w={last_w:.4f} top3_pos={top3_pos.indices.tolist()}")
     else:
         # Use FMHA kernel for longer sequences (padding effect is negligible)
         from dsv4.kernels.attention.fmha_multitile_op import fmha_multitile_decode_raw