diff --git a/src/config.py b/src/config.py
index 3170cd1..5b19eb6 100644
--- a/src/config.py
+++ b/src/config.py
@@ -101,6 +101,7 @@ class Diagnostic(BaseModel):
     est_cma: Optional[bool] = None
     est_cms: Optional[bool] = None
     niveau_severite: Optional[str] = None  # "leger" | "modere" | "severe" | "non_evalue"
+    source: Optional[str] = None  # "trackare" | "edsnlp" | "regex" | "llm_das"
 
 
 class ActeCCAM(BaseModel):
diff --git a/src/medical/cim10_dict.py b/src/medical/cim10_dict.py
index 33d15ec..bf2d83e 100644
--- a/src/medical/cim10_dict.py
+++ b/src/medical/cim10_dict.py
@@ -212,6 +212,25 @@ def validate_code(code: str) -> tuple[bool, str]:
     return False, ""
 
 
+def fallback_parent_code(code: str) -> str | None:
+    """Tente de corriger un code invalide en remontant au code parent.
+
+    Le LLM hallucine souvent des sous-codes (.8, .9) sur des codes
+    standalone à 3 caractères (ex: D71.9 → D71, R69.8 → R69).
+
+    Returns:
+        Le code parent valide, ou None si aucun fallback trouvé.
+    """
+    normalized = normalize_code(code)
+    # Extraire le code parent (3 caractères avant le point)
+    if "." in normalized:
+        parent = normalized.split(".")[0]
+        is_valid, _ = validate_code(parent)
+        if is_valid:
+            return parent
+    return None
+
+
 def reset_cache() -> None:
     """Réinitialise les caches (utile pour les tests)."""
     global _dict_cache, _normalized_cache
diff --git a/src/medical/cim10_extractor.py b/src/medical/cim10_extractor.py
index c8816e8..8c9da1e 100644
--- a/src/medical/cim10_extractor.py
+++ b/src/medical/cim10_extractor.py
@@ -205,6 +205,7 @@ def _extract_das_llm(text: str, dossier: DossierMedical) -> None:
                 texte=texte,
                 cim10_suggestion=code,
                 justification=das.get("justification"),
+                source="llm_das",
             ))
             added += 1
 
@@ -297,6 +298,7 @@ def _extract_diagnostics(
         d = Diagnostic(
             texte=texte,
             cim10_suggestion=diag.get("code_cim10"),
+            source="trackare",
         )
         if diag.get("type", "").lower() == "principal":
             dossier.diagnostic_principal = d
@@ -331,6 +333,7 @@ def _extract_diagnostics(
             code, texte = next(iter(edsnlp_codes.items()))
             dossier.diagnostic_principal = Diagnostic(
                 texte=texte.capitalize(), cim10_suggestion=code,
+                source="edsnlp",
             )
 
     # Diagnostics associés depuis le texte (regex)
@@ -356,6 +359,7 @@ def _extract_diagnostics(
                 dossier.diagnostics_associes.append(Diagnostic(
                     texte=texte,
                     cim10_suggestion=ent.code,
+                    source="edsnlp",
                 ))
                 existing_codes.add(ent.code)
 
@@ -370,7 +374,7 @@ def _find_diagnostic_principal(text_lower: str, conclusion: str) -> Diagnostic |
     # Chercher dans la conclusion d'abord via CIM10_MAP (domain override)
     for terme, code in CIM10_MAP.items():
         if normalize_text(terme) in conclusion_norm:
-            return Diagnostic(texte=terme.capitalize(), cim10_suggestion=code)
+            return Diagnostic(texte=terme.capitalize(), cim10_suggestion=code, source="regex")
 
     text_norm = normalize_text(text_lower)
 
@@ -385,7 +389,7 @@ def _find_diagnostic_principal(text_lower: str, conclusion: str) -> Diagnostic |
         if m:
             matched = m.group(0)
             code = _lookup_cim10(matched)
-            return Diagnostic(texte=matched.capitalize(), cim10_suggestion=code)
+            return Diagnostic(texte=matched.capitalize(), cim10_suggestion=code, source="regex")
 
     return None
 
@@ -444,7 +448,7 @@ def _find_diagnostics_associes(
     # Patterns DAS
     for pat, label, code in _DAS_PATTERNS:
         if re.search(pat, text_norm) and code not in existing_codes:
-            das.append(Diagnostic(texte=label, cim10_suggestion=code))
+            das.append(Diagnostic(texte=label, cim10_suggestion=code, source="regex"))
             existing_codes.add(code)
 
     # Obésité (IMC >= 30) — pattern spécial avec extraction de valeur
@@ -452,7 +456,7 @@ def _find_diagnostics_associes(
     if m:
         imc_val = float(m.group(1).replace(",", "."))
         if imc_val >= 30 and "E66.0" not in existing_codes:
-            das.append(Diagnostic(texte=f"Obésité (IMC {imc_val})", cim10_suggestion="E66.0"))
+            das.append(Diagnostic(texte=f"Obésité (IMC {imc_val})", cim10_suggestion="E66.0", source="regex"))
             existing_codes.add("E66.0")
 
     return das
diff --git a/src/medical/das_filter.py b/src/medical/das_filter.py
index 5023bc1..af61e4b 100644
--- a/src/medical/das_filter.py
+++ b/src/medical/das_filter.py
@@ -83,6 +83,20 @@ def is_valid_diagnostic_text(text: str) -> bool:
     if re.match(r'^(Dans |La |Le |Les |Au |Aux )', t) and len(t) < 30:
         return False
 
+    # 12. En-têtes de systèmes anatomiques (catégories sans pathologie)
+    _ANATOMICAL_HEADERS = {
+        "musculaire", "squelettique", "cardiovasculaire", "pulmonaire",
+        "neurologique", "digestif", "digestive", "hépatique", "rénal",
+        "rénale", "urinaire", "cutané", "cutanée", "articulaire",
+        "osseux", "osseuse", "gastrique", "intestinal", "intestinale",
+        "cérébral", "thoracique", "abdominal", "abdominale",
+    }
+    if len(words) == 1 and t.lower() in _ANATOMICAL_HEADERS:
+        return False
+    # Catégorie + description vague : "Musculaire - masse musculaire"
+    if re.match(r'^[A-ZÀ-Ú][a-zà-ÿ]+ - (masse|zone|région|état|bilan)', t, re.IGNORECASE):
+        return False
+
     return True
 
 
diff --git a/src/medical/rag_search.py b/src/medical/rag_search.py
index 1a86b01..ff9cc61 100644
--- a/src/medical/rag_search.py
+++ b/src/medical/rag_search.py
@@ -10,7 +10,7 @@ from ..config import (
     OLLAMA_CACHE_PATH, OLLAMA_MAX_PARALLEL, OLLAMA_MODEL,
     EMBEDDING_MODEL, RERANKER_MODEL,
 )
-from .cim10_dict import normalize_code, validate_code as cim10_validate
+from .cim10_dict import normalize_code, validate_code as cim10_validate, fallback_parent_code
 from .cim10_extractor import BIO_NORMALS
 from .ccam_dict import validate_code as ccam_validate
 from .ollama_client import call_ollama, parse_json_response
@@ -478,10 +478,19 @@ def _apply_llm_result_diagnostic(diagnostic: Diagnostic, llm_result: dict) -> No
         if is_valid:
             diagnostic.cim10_suggestion = code
         else:
-            logger.warning(
-                "RAG : code Ollama %s invalide pour « %s », code ignoré",
-                code, diagnostic.texte,
-            )
+            # Tenter fallback vers le code parent (D71.9 → D71)
+            parent = fallback_parent_code(code)
+            if parent:
+                logger.info(
+                    "RAG : code Ollama %s invalide → fallback parent %s pour « %s »",
+                    code, parent, diagnostic.texte,
+                )
+                diagnostic.cim10_suggestion = parent
+            else:
+                logger.warning(
+                    "RAG : code Ollama %s invalide pour « %s », code ignoré",
+                    code, diagnostic.texte,
+                )
     if confidence in ("high", "medium", "low"):
         diagnostic.cim10_confidence = confidence
     if justification:
diff --git a/tests/test_das_filter.py b/tests/test_das_filter.py
index 009a1d1..9c0913b 100644
--- a/tests/test_das_filter.py
+++ b/tests/test_das_filter.py
@@ -161,6 +161,29 @@ class TestIsValidDiagnosticText:
         """Un fragment long commençant par 'Dans' peut être légitime."""
         assert is_valid_diagnostic_text("Dans le cadre d'une insuffisance rénale chronique terminale")
 
+    # --- Règle 12 : en-têtes anatomiques / catégories vagues ---
+    def test_reject_musculaire_alone(self):
+        assert not is_valid_diagnostic_text("Musculaire")
+
+    def test_reject_musculaire_masse(self):
+        assert not is_valid_diagnostic_text("Musculaire - masse musculaire")
+
+    def test_reject_digestif_alone(self):
+        assert not is_valid_diagnostic_text("Digestif")
+
+    def test_reject_hepatique_alone(self):
+        assert not is_valid_diagnostic_text("Hépatique")
+
+    def test_reject_category_bilan(self):
+        assert not is_valid_diagnostic_text("Rénal - bilan rénal")
+
+    def test_accept_real_diagnostic_musculaire(self):
+        """Un vrai diagnostic contenant 'musculaire' est accepté."""
+        assert is_valid_diagnostic_text("Dystrophie musculaire de Duchenne")
+
+    def test_accept_real_diagnostic_hepatique(self):
+        assert is_valid_diagnostic_text("Insuffisance hépatique aiguë")
+
 
 class TestCorrectKnownMiscodes:
     """Tests pour la correction des codes systématiquement mal attribués."""
diff --git a/tests/test_rag.py b/tests/test_rag.py
index 184ec6b..b4b70d9 100644
--- a/tests/test_rag.py
+++ b/tests/test_rag.py
@@ -504,7 +504,7 @@ class TestRAGSearchMocked:
             {"document": "cim10", "page": 496, "code": "K85", "extrait": "K85", "score": 0.9},
         ]
         mock_llm = {
-            "code": "X99.99",  # code invalide
+            "code": "QQ9.99",  # code invalide (pas de parent valide)
             "confidence": "high",
             "justification": "Hallucination",
         }
@@ -516,6 +516,26 @@ class TestRAGSearchMocked:
         # Le code original est conservé (pas remplacé par le code invalide)
         assert diag.cim10_suggestion == "K85.9"
 
+    def test_enrich_diagnostic_fallback_parent_code(self):
+        """Un code invalide D71.9 est corrigé en D71 (code parent standalone)."""
+        from src.medical.rag_search import enrich_diagnostic
+
+        diag = Diagnostic(texte="Anomalie des leucocytes")
+        mock_sources = [
+            {"document": "cim10", "page": 100, "code": "D71", "extrait": "D71", "score": 0.9},
+        ]
+        mock_llm = {
+            "code": "D71.9",  # invalide : D71 est standalone
+            "confidence": "high",
+            "justification": "Anomalie leucocytaire",
+        }
+
+        with patch("src.medical.rag_search.search_similar", return_value=mock_sources), \
+             patch("src.medical.rag_search._call_ollama", return_value=mock_llm):
+            enrich_diagnostic(diag, {"sexe": "M", "age": 60})
+
+        assert diag.cim10_suggestion == "D71"
+
     def test_enrich_diagnostic_normalizes_code(self):
         """Un code Ollama sans point est normalisé (K851 → K85.1)."""
         from src.medical.rag_search import enrich_diagnostic
@@ -668,6 +688,36 @@ class TestValidateCodeCIM10:
         assert is_valid is True
 
 
+class TestFallbackParentCode:
+    def test_d71_9_to_d71(self):
+        """D71.9 (invalide) → D71 (standalone valide)."""
+        from src.medical.cim10_dict import fallback_parent_code
+        assert fallback_parent_code("D71.9") == "D71"
+
+    def test_r69_8_to_r69(self):
+        """R69.8 (invalide) → R69 (standalone valide)."""
+        from src.medical.cim10_dict import fallback_parent_code
+        assert fallback_parent_code("R69.8") == "R69"
+
+    def test_valid_code_no_fallback(self):
+        """Un code déjà valide ne devrait pas matcher (parent aussi valide)."""
+        from src.medical.cim10_dict import fallback_parent_code
+        # K85.1 est valide, donc on ne devrait pas appeler fallback
+        # Mais si on l'appelle, K85 est aussi valide → retourne K85
+        result = fallback_parent_code("K85.1")
+        assert result == "K85"  # le parent est valide
+
+    def test_truly_invalid_no_fallback(self):
+        """Un code sans parent valide retourne None."""
+        from src.medical.cim10_dict import fallback_parent_code
+        assert fallback_parent_code("QQ9.99") is None
+
+    def test_three_char_code_no_fallback(self):
+        """Un code 3 caractères sans point ne peut pas remonter."""
+        from src.medical.cim10_dict import fallback_parent_code
+        assert fallback_parent_code("QQ9") is None
+
+
 class TestValidateCIM10PostProcessing:
     def test_hallucination_rejected(self):
         """Les codes hallucination (Aucun, N/A...) sont rejetés."""