diff --git a/tests/unit/test_d5c_multitile.py b/tests/unit/test_d5c_multitile.py
index c0e01bcc..06581c20 100644
--- a/tests/unit/test_d5c_multitile.py
+++ b/tests/unit/test_d5c_multitile.py
@@ -129,6 +129,37 @@ def test_d5c_multitile():
     attn_sink_val = 0.5
     attn_sink = torch.tensor([attn_sink_val], dtype=torch.float32, device='cuda')
 
+    # Direct test of segment 0 (verify run_segment isn't the issue)
+    print('--- Direct segment 0 test ---')
+    k_seg0 = k_combined[0:seg_size]
+    v_seg0 = v_combined[0:seg_size]
+    c0 = torch.zeros(m, hd, 1, dtype=torch.bfloat16, device='cuda')
+    lse0 = torch.zeros(m, 1, 1, dtype=torch.float32, device='cuda')
+    rs0 = torch.zeros(m, 1, 1, dtype=torch.float32, device='cuda')
+    mQ0 = to_cute(q.unsqueeze(-1))
+    mK0 = to_cute(k_seg0.unsqueeze(-1))
+    mV0 = to_cute(v_seg0.unsqueeze(-1))
+    mC0 = to_cute(c0); mLSE0 = to_cute(lse0); mRS0 = to_cute(rs0)
+    mSB0 = to_cute(attn_sink)
+    # Compile FRESH for segment 0
+    k0 = FmhaKernel(head_dim=hd, s_k=seg_size, normalize=False,
+                    apply_swa_mask=True, is_causal=False, n_comp=n_comp, apply_sink_bias=True)
+    comp0 = cute.compile(k0, mQ0, mK0, mV0, mC0, stream, mLSE0,
+                         swa_len=swa_len, sink_bias=mSB0, row_sums=mRS0)
+    comp0(mQ0, mK0, mV0, mC0, stream, mLSE0,
+          swa_len=swa_len, sink_bias=mSB0, row_sums=mRS0)
+    torch.cuda.synchronize()
+    ok0 = c0[:, :, 0].float() / rs0[:, 0, 0].float().unsqueeze(1).clamp(min=1e-30)
+    # Per-segment reference for segment 0
+    scores0_ref = q.float() @ k_seg0.float().T * scale
+    scores0_ref[:, n_comp:] += attn_sink_val
+    if swa_len < (seg_size - n_comp):  # 100 < 32? No
+        scores0_ref[:, n_comp + swa_len:] = float('-inf')
+    ref0 = (torch.softmax(scores0_ref, dim=-1) @ v_seg0.float()).to(torch.bfloat16)
+    cos0 = torch.nn.functional.cosine_similarity(ok0.flatten().unsqueeze(0), ref0.flatten().unsqueeze(0).float()).item()
+    print(f'Seg0 direct: cos {cos0:.6f}  LSE_kern={lse0[0,0,0].item():.4f}  LSE_ref={torch.logsumexp(scores0_ref, dim=-1)[0].item():.4f}')
+    print()
+
     # Reference
     ref = reference_combined_attention(
         q, k_comp, v_comp, k_swa, v_swa,