diff --git a/core/execution/observe_reason_act.py b/core/execution/observe_reason_act.py
index cf9e07614..893e0d84a 100644
--- a/core/execution/observe_reason_act.py
+++ b/core/execution/observe_reason_act.py
@@ -925,6 +925,16 @@ Règles:
                 logger.error(f"❌ [ORA/click] Impossible de localiser '{target_text}' — aucune méthode n'a fonctionné")
                 return False
 
+        # --- Vérification pré-action : est-ce le bon élément ? ---
+        if target_text and method_used not in ('template',) and MSS_AVAILABLE and PIL_AVAILABLE:
+            try:
+                pre_check = self._verify_pre_click(x, y, target_text, target_desc)
+                if not pre_check:
+                    print(f"⛔ [ORA/pre-check] L'élément à ({x}, {y}) ne correspond PAS à '{target_text}' — abandon du clic")
+                    return False
+            except Exception as e:
+                print(f"⚠️ [ORA/pre-check] Erreur vérification: {e}")
+
         print(f"🖱️ [ORA/click] {decision.value} à ({x}, {y}) via {method_used}")
 
         if decision.value == 'double':
@@ -1079,6 +1089,55 @@ Règles:
             pass
         return ''
 
+    def _verify_pre_click(self, x: int, y: int, target_text: str, target_desc: str = "") -> bool:
+        """Vérifie que l'élément à la position (x,y) correspond au target AVANT de cliquer.
+
+        Fait un crop 200x100 autour de (x,y), envoie au VLM avec la question
+        "est-ce que c'est bien {target} ?"
+        """
+        try:
+            import requests as _requests
+
+            with mss_lib.mss() as sct:
+                mon = sct.monitors[0]
+                grab = sct.grab(mon)
+                screen = Image.frombytes('RGB', grab.size, grab.bgra, 'raw', 'BGRX')
+
+            # Crop 200x100 autour du point de clic
+            crop_w, crop_h = 200, 100
+            left = max(0, x - crop_w // 2)
+            top = max(0, y - crop_h // 2)
+            right = min(screen.width, left + crop_w)
+            bottom = min(screen.height, top + crop_h)
+            crop = screen.crop((left, top, right, bottom))
+
+            import io as _io
+            buffer = _io.BytesIO()
+            crop.save(buffer, format='JPEG', quality=70)
+            crop_b64 = base64.b64encode(buffer.getvalue()).decode('utf-8')
+
+            label = target_desc or target_text
+            ollama_url = os.environ.get("OLLAMA_URL", "http://localhost:11434")
+            resp = _requests.post(f"{ollama_url}/api/generate", json={
+                "model": "qwen2.5vl:3b",
+                "prompt": f"Is this UI element '{label}'? Answer only YES or NO.",
+                "images": [crop_b64],
+                "stream": False,
+                "options": {"temperature": 0.1, "num_predict": 5}
+            }, timeout=15)
+
+            if resp.status_code == 200:
+                answer = resp.json().get("response", "").strip().upper()
+                is_match = "YES" in answer
+                print(f"🔍 [ORA/pre-check] '{label}' → {answer} → {'✅' if is_match else '❌'}")
+                return is_match
+
+            return True  # En cas d'erreur HTTP, on laisse passer
+
+        except Exception as e:
+            print(f"⚠️ [ORA/pre-check] Erreur: {e}")
+            return True  # En cas d'erreur, on laisse passer
+
     def _phash_distance(self, hash1: Any, hash2: Any) -> int:
         """Distance de Hamming entre deux pHash. Retourne 999 si non calculable."""
         if hash1 is None or hash2 is None: