diff --git a/klausur-service/backend/cv_vocab_pipeline.py b/klausur-service/backend/cv_vocab_pipeline.py
index d56323d..3c4e304 100644
--- a/klausur-service/backend/cv_vocab_pipeline.py
+++ b/klausur-service/backend/cv_vocab_pipeline.py
@@ -3103,8 +3103,9 @@ def _ocr_single_cell(
             'ocr_engine': 'word_lookup',
         }
 
-    # Use pre-assigned words (exclusive per column) if provided
+    # --- PRIMARY: Word-lookup from full-page Tesseract ---
     words = preassigned_words if preassigned_words is not None else []
+    used_engine = 'word_lookup'
 
     if words:
         # Use row height as Y-tolerance so all words within a single row
@@ -3112,12 +3113,36 @@ def _ocr_single_cell(
         # across two lines due to slight vertical offset).
         y_tol = max(15, row.height)
         text = _words_to_reading_order_text(words, y_tolerance_px=y_tol)
+        avg_conf = round(sum(w['conf'] for w in words) / len(words), 1)
     else:
         text = ''
+        avg_conf = 0.0
 
-    avg_conf = 0.0
-    if words:
-        avg_conf = round(sum(w['conf'] for w in words) / len(words), 1)
+    # --- FALLBACK: Cell-OCR for empty cells ---
+    # Full-page Tesseract can miss small or isolated words (e.g. "Ei").
+    # Re-run OCR on the cell crop to catch what word-lookup missed.
+    if not text.strip() and cell_w > 0 and cell_h > 0:
+        cell_region = PageRegion(
+            type=col.type,
+            x=cell_x, y=cell_y,
+            width=cell_w, height=cell_h,
+        )
+        if use_rapid and img_bgr is not None:
+            fallback_words = ocr_region_rapid(img_bgr, cell_region)
+        else:
+            cell_lang = lang_map.get(col.type, lang)
+            fallback_words = ocr_region(ocr_img, cell_region, lang=cell_lang, psm=6)
+
+        if fallback_words:
+            fb_avg_h = sum(w['height'] for w in fallback_words) / len(fallback_words)
+            fb_y_tol = max(10, int(fb_avg_h * 0.5))
+            fb_text = _words_to_reading_order_text(fallback_words, y_tolerance_px=fb_y_tol)
+            if fb_text.strip():
+                text = fb_text
+                avg_conf = round(
+                    sum(w['conf'] for w in fallback_words) / len(fallback_words), 1
+                )
+                used_engine = 'cell_ocr_fallback'
 
     return {
         'cell_id': f"R{row_idx:02d}_C{col_idx}",
@@ -3133,7 +3158,7 @@ def _ocr_single_cell(
             'w': round(cell_w / img_w * 100, 2),
             'h': round(cell_h / img_h * 100, 2),
         },
-        'ocr_engine': 'word_lookup',
+        'ocr_engine': used_engine,
     }