--- a/common/speculative.cpp	2026-10-13 05:29:52.854135340 -0400
+++ b/common/speculative.cpp	2026-10-13 05:29:52.967135794 -0400
@@ -2132,6 +2132,7 @@
     // Row 0 corresponds to the sampled token, row N to the Nth accepted draft token.
     std::vector<std::vector<float>> verify_h;
     std::vector<int32_t> verify_h_rows;
+    std::vector<llama_pos> verify_pos_first; // position of verify_h row 0
 
     std::vector<int>                i_last;
     std::vector<std::vector<float>> chain_h;
@@ -2221,6 +2222,7 @@
 
         verify_h.assign(n_seq, {});
         verify_h_rows.assign(n_seq, 0);
+        verify_pos_first.assign(n_seq, -1);
 
         deferred_tok.assign(n_seq, {});
         deferred_pos.assign(n_seq, {});
@@ -2443,7 +2445,8 @@
             }
 
             const int32_t n_rows = i_batch_end[seq_id] - i_batch_beg[seq_id] + 1;
-            verify_h_rows[seq_id] = n_rows;
+            verify_h_rows[seq_id]    = n_rows;
+            verify_pos_first[seq_id] = batch_in.pos[i_batch_beg[seq_id]];
             verify_h[seq_id].resize((size_t) n_rows * n_embd);
 
             for (int32_t i = 0; i < n_rows; ++i) {
@@ -2627,6 +2630,11 @@
         const int32_t i_h = std::min<int32_t>(n_accepted, n_rows - 1);
         const size_t row_bytes = (size_t) n_embd * sizeof(float);
         std::memcpy(pending_h[seq_id].data(), verify_h[seq_id].data() + (size_t) i_h * n_embd, row_bytes);
+
+        // the rejected rows of the verification batch must never reach ctx_dft: draft() skips them
+        // (add_deferred stops at n_past), but a flush decodes every deferred row, and a flush happens
+        // whenever a batch without this seq is too big to defer (a fair prompt step) or the slot resets
+        drop_deferred_from(seq_id, verify_pos_first[seq_id] + i_h + 1);
     }
 };