diff --git a/klausur-service/backend/cv_vocab_pipeline.py b/klausur-service/backend/cv_vocab_pipeline.py
index 4dee8c5..a53d3af 100644
--- a/klausur-service/backend/cv_vocab_pipeline.py
+++ b/klausur-service/backend/cv_vocab_pipeline.py
@@ -3093,43 +3093,56 @@ def _assign_row_words_to_columns(
 
     left_x = row.x  # content ROI left (absolute)
 
-    # Pre-compute column bounds and centers in relative coordinates
-    col_bounds_rel = []  # (left, right, center) per column
-    for col in columns:
+    # Build non-overlapping column assignment ranges using midpoints.
+    # For adjacent columns, the boundary is the midpoint between them.
+    # This prevents words near column borders from being assigned to
+    # the wrong column (e.g. "We" at the start of an example sentence
+    # being stolen by the preceding DE column).
+    n = len(columns)
+    col_ranges_rel = []  # (assign_left, assign_right) per column
+    for ci, col in enumerate(columns):
         col_left_rel = col.x - left_x
         col_right_rel = col_left_rel + col.width
-        col_center_rel = col_left_rel + col.width / 2
-        col_bounds_rel.append((col_left_rel, col_right_rel, col_center_rel))
 
-    # Padding: allow words slightly outside column bounds (e.g. due to
-    # imprecise column detection).  Use 15% of average column width.
-    avg_col_w = sum(c.width for c in columns) / len(columns) if columns else 100
-    pad = avg_col_w * 0.15
+        # Left boundary: midpoint to previous column, or 0
+        if ci == 0:
+            assign_left = 0
+        else:
+            prev_right = columns[ci - 1].x - left_x + columns[ci - 1].width
+            assign_left = (prev_right + col_left_rel) / 2
+
+        # Right boundary: midpoint to next column, or infinity (row width)
+        if ci == n - 1:
+            assign_right = row.width + 100  # generous for last column
+        else:
+            next_left = columns[ci + 1].x - left_x
+            assign_right = (col_right_rel + next_left) / 2
+
+        col_ranges_rel.append((assign_left, assign_right))
 
     for w in row.words:
         w_center_x = w['left'] + w['width'] / 2
 
-        # Pass 1: containment check (word center within column bounds + pad)
-        contained_col = -1
-        for ci, (cl, cr, _) in enumerate(col_bounds_rel):
-            if (cl - pad) <= w_center_x <= (cr + pad):
-                contained_col = ci
+        # Find which column range contains this word
+        assigned = False
+        for ci, (al, ar) in enumerate(col_ranges_rel):
+            if al <= w_center_x < ar:
+                result[ci].append(w)
+                assigned = True
                 break
 
-        if contained_col >= 0:
-            result[contained_col].append(w)
-            continue
-
-        # Pass 2: nearest center fallback
-        best_col = 0
-        best_dist = abs(w_center_x - col_bounds_rel[0][2])
-        for ci in range(1, len(columns)):
-            dist = abs(w_center_x - col_bounds_rel[ci][2])
-            if dist < best_dist:
-                best_dist = dist
-                best_col = ci
-
-        result[best_col].append(w)
+        if not assigned:
+            # Fallback: nearest column center
+            best_col = 0
+            col_left_0 = columns[0].x - left_x
+            best_dist = abs(w_center_x - (col_left_0 + columns[0].width / 2))
+            for ci in range(1, n):
+                col_left = columns[ci].x - left_x
+                dist = abs(w_center_x - (col_left + columns[ci].width / 2))
+                if dist < best_dist:
+                    best_dist = dist
+                    best_col = ci
+            result[best_col].append(w)
 
     return result