diff --git a/core/execution/observe_reason_act.py b/core/execution/observe_reason_act.py
index e838ced04..ee4616683 100644
--- a/core/execution/observe_reason_act.py
+++ b/core/execution/observe_reason_act.py
@@ -1235,25 +1235,28 @@ Règles:
             pre = self.observe()
 
             # --- 1b. Réflexe Check : popup/dialogue inattendu ? ---
-            # Si un dialogue connu est détecté (Enregistrer sous, Oui/Non, OK, etc.)
-            # Léa le gère immédiatement AVANT de continuer le workflow.
-            try:
-                from core.execution.input_handler import check_screen_for_patterns, handle_detected_pattern
-                _reflex_pattern = check_screen_for_patterns()
-                if _reflex_pattern:
-                    _reflex_name = _reflex_pattern.get('pattern', '?')
-                    _reflex_target = _reflex_pattern.get('target', '?')
-                    print(f"🧠 [ORA/réflexe] Pattern détecté: '{_reflex_name}' → clic '{_reflex_target}'")
-                    _handled = handle_detected_pattern(_reflex_pattern)
-                    if _handled:
-                        print(f"✅ [ORA/réflexe] Dialogue '{_reflex_name}' géré automatiquement")
-                        time.sleep(0.5)
-                        # Re-observer après avoir géré le dialogue
-                        pre = self.observe()
-                    else:
-                        print(f"⚠️ [ORA/réflexe] Pattern '{_reflex_name}' détecté mais non géré")
-            except Exception as _reflex_err:
-                print(f"⚠️ [ORA/réflexe] Erreur: {_reflex_err}")
+            # Déclenché UNIQUEMENT si le pHash a changé de manière inattendue
+            # (= un popup est probablement apparu). Sinon → 0ms, pas d'OCR.
+            if i > 0 and hasattr(self, '_last_post_phash') and self._last_post_phash:
+                _phash_distance = self._phash_distance(pre.phash, self._last_post_phash)
+                if _phash_distance > 10:  # Changement significatif inattendu
+                    print(f"🧠 [ORA/réflexe] pHash changé (distance={_phash_distance}) → vérification popup")
+                    try:
+                        from core.execution.input_handler import check_screen_for_patterns, handle_detected_pattern
+                        _reflex_pattern = check_screen_for_patterns()
+                        if _reflex_pattern:
+                            _reflex_name = _reflex_pattern.get('pattern', '?')
+                            _reflex_target = _reflex_pattern.get('target', '?')
+                            print(f"🧠 [ORA/réflexe] Pattern détecté: '{_reflex_name}' → clic '{_reflex_target}'")
+                            _handled = handle_detected_pattern(_reflex_pattern)
+                            if _handled:
+                                print(f"✅ [ORA/réflexe] Dialogue '{_reflex_name}' géré automatiquement")
+                                time.sleep(0.5)
+                                pre = self.observe()
+                            else:
+                                print(f"⚠️ [ORA/réflexe] Pattern '{_reflex_name}' détecté mais non géré")
+                    except Exception as _reflex_err:
+                        print(f"⚠️ [ORA/réflexe] Erreur: {_reflex_err}")
 
             # --- 2. Raisonner : construire la Decision ---
             decision = self.reason_workflow_step(step, pre)
@@ -1307,6 +1310,8 @@ Règles:
 
             # --- 4. Observer l'état post-action ---
             post = self.observe()
+            # Stocker le pHash post-action pour le réflexe check du step suivant
+            self._last_post_phash = post.phash
 
             # --- 5. Vérifier ---
             verification = self.verify(pre, post, decision)
diff --git a/core/grounding/server.py b/core/grounding/server.py
index 7d757dbd2..532f7b5e4 100644
--- a/core/grounding/server.py
+++ b/core/grounding/server.py
@@ -360,7 +360,7 @@ def ground(req: GroundRequest):
         # Inference
         t0 = time.time()
         with torch.no_grad():
-            gen = _model.generate(**inputs, max_new_tokens=256)
+            gen = _model.generate(**inputs, max_new_tokens=64)
         infer_ms = (time.time() - t0) * 1000
 
         # Decoder