diff --git a/backend/app/config.py b/backend/app/config.py
index 9822768..e7b26d4 100644
--- a/backend/app/config.py
+++ b/backend/app/config.py
@@ -39,9 +39,8 @@ class Settings(BaseSettings):
     RATE_LIMIT_AUTH: int = 10
     RATE_LIMIT_VOTE: int = 30
 
-    # AI (Claude — substitut Qwen en attendant le déploiement local)
-    ANTHROPIC_API_KEY: str = ""
-    ANTHROPIC_MODEL: str = "claude-haiku-4-5-20251001"
+    # AI — Qwen3.6 (MacStudio) endpoint, branché plus tard
+    QWEN_API_URL: str = ""
 
     # Blockchain cache
     BLOCKCHAIN_CACHE_TTL_SECONDS: int = 3600
diff --git a/backend/app/routers/qualify.py b/backend/app/routers/qualify.py
index aab005c..a01dca5 100644
--- a/backend/app/routers/qualify.py
+++ b/backend/app/routers/qualify.py
@@ -25,7 +25,7 @@ from app.services.qualify_ai_service import (
     AIMessage,
     AIQuestion,
     AIQualifyResult,
-    ai_frame_async,
+    ai_frame,
 )
 
 router = APIRouter()
@@ -127,7 +127,7 @@ async def ai_chat(payload: AIChatRequest) -> AIChatResponse:
         context=payload.context,
         messages=[AIMessage(role=m.role, content=m.content) for m in payload.messages],
     )
-    resp = await ai_frame_async(req)
+    resp = ai_frame(req)
 
     return AIChatResponse(
         done=resp.done,
diff --git a/backend/app/services/qualify_ai_service.py b/backend/app/services/qualify_ai_service.py
index c86b2c8..dab7b20 100644
--- a/backend/app/services/qualify_ai_service.py
+++ b/backend/app/services/qualify_ai_service.py
@@ -3,23 +3,14 @@
 Orchestrates a 2-round conversation that clarifies reversibility and urgency
 before producing a final QualificationResult.
 
-Two entry points:
-  ai_frame()       — synchronous, rule-based only. Used by tests.
-  ai_frame_async() — async, enriched by Claude API (or Qwen3.6 later).
-                     Falls back to ai_frame() if ANTHROPIC_API_KEY is unset.
-
-The interface (AIFrameRequest / AIFrameResponse) is stable; the underlying
-engine is swappable (Qwen3.6 MacStudio planned to replace Claude calls).
+Rule-based stub — Qwen3.6 (MacStudio) calls will replace ai_frame() internals
+once the local endpoint is available. The interface is stable.
 """
 
 from __future__ import annotations
 
-import json
-import logging
 from dataclasses import dataclass
 
-logger = logging.getLogger(__name__)
-
 
 # ---------------------------------------------------------------------------
 # Schemas (dataclasses — no Pydantic dependency in the engine layer)
@@ -74,7 +65,7 @@ class AIFrameResponse:
 
 
 # ---------------------------------------------------------------------------
-# Standard clarifying questions (fallback when no context or API unavailable)
+# Standard clarifying questions
 # ---------------------------------------------------------------------------
 
 
@@ -101,12 +92,12 @@ _CLARIFYING_QUESTIONS: list[AIQuestion] = [
 
 
 # ---------------------------------------------------------------------------
-# Synchronous rule-based engine (used directly by tests)
+# Core function
 # ---------------------------------------------------------------------------
 
 
 def ai_frame(request: AIFrameRequest) -> AIFrameResponse:
-    """Run one round of rule-based AI framing.
+    """Run one round of AI framing.
 
     Round 1 (messages=[])  → return 2 clarifying questions, done=False
     Round 2 (messages set) → parse answers, qualify, return result, done=True
@@ -134,135 +125,7 @@ def ai_frame(request: AIFrameRequest) -> AIFrameResponse:
 
 
 # ---------------------------------------------------------------------------
-# Async Claude-enriched entry point (used by the router)
-# ---------------------------------------------------------------------------
-
-
-async def ai_frame_async(request: AIFrameRequest) -> AIFrameResponse:
-    """Async entry point: rule engine + Claude enrichment.
-
-    Falls back to ai_frame() if ANTHROPIC_API_KEY is not configured.
-    Qwen3.6 (MacStudio) will replace Claude calls when available.
-    """
-    base = ai_frame(request)
-
-    client = _get_claude_client()
-    if client is None:
-        return base
-
-    messages = request.messages or []
-
-    if not messages:
-        # Round 1: context-aware questions
-        enriched_questions = await _claude_questions(client, request.context, base.questions)
-        return AIFrameResponse(done=False, questions=enriched_questions, result=None, explanation=None)
-    else:
-        # Round 2: enriched explanation
-        enriched_explanation = await _claude_explanation(client, request, base)
-        return AIFrameResponse(
-            done=True,
-            questions=[],
-            result=base.result,
-            explanation=enriched_explanation or base.explanation,
-        )
-
-
-# ---------------------------------------------------------------------------
-# Claude helpers
-# ---------------------------------------------------------------------------
-
-
-def _get_claude_client():
-    """Return an AsyncAnthropic client if ANTHROPIC_API_KEY is set, else None."""
-    try:
-        from app.config import settings
-        if not settings.ANTHROPIC_API_KEY:
-            return None
-        import anthropic
-        return anthropic.AsyncAnthropic(api_key=settings.ANTHROPIC_API_KEY)
-    except Exception:
-        return None
-
-
-async def _claude_questions(client, context: str | None, fallback: list[AIQuestion]) -> list[AIQuestion]:
-    """Generate context-aware clarifying questions. Falls back to standard questions on error."""
-    if not context:
-        return fallback
-
-    from app.config import settings
-    prompt = f"""Tu assistes à la qualification d'une décision collective dans la communauté Duniter/G1 (monnaie libre).
-
-Contexte de la décision : {context}
-
-Génère 2 questions de clarification courtes et précises pour mieux qualifier cette décision.
-Chaque question doit avoir exactement 3 options de réponse.
-Les questions doivent porter sur la réversibilité et l'urgence, adaptées au contexte.
-
-Réponds UNIQUEMENT en JSON valide, sans texte avant ni après :
-[
-  {{"id": "reversibility", "text": "...", "options": ["option1", "option2", "option3"]}},
-  {{"id": "urgency", "text": "...", "options": ["option1", "option2", "option3"]}}
-]"""
-
-    try:
-        message = await client.messages.create(
-            model=settings.ANTHROPIC_MODEL,
-            max_tokens=512,
-            messages=[{"role": "user", "content": prompt}],
-        )
-        raw = message.content[0].text.strip()
-        start = raw.find("[")
-        end = raw.rfind("]") + 1
-        data = json.loads(raw[start:end])
-        return [AIQuestion(id=q["id"], text=q["text"], options=q["options"]) for q in data]
-    except Exception as exc:
-        logger.warning("Claude question generation failed: %s — using fallback", exc)
-        return fallback
-
-
-async def _claude_explanation(client, request: AIFrameRequest, base: AIFrameResponse) -> str | None:
-    """Generate a contextual explanation for the qualification result."""
-    if base.result is None:
-        return None
-
-    from app.config import settings
-    answers = _parse_answers(request.messages or [])
-    rev = answers.get("reversibility", "non précisé")
-    urg = answers.get("urgency", "non précisé")
-    context_line = f"\nContexte : {request.context}" if request.context else ""
-
-    prompt = f"""Tu qualifies une décision collective pour la communauté Duniter/G1 (monnaie libre).{context_line}
-
-Paramètres :
-- Personnes concernées : {request.affected_count or 'non précisé'}
-- Dans le cadre d'un mandat : {'oui' if request.within_mandate else 'non'}
-- Décision structurante (on-chain) : {'oui' if request.is_structural else 'non'}
-- Réversibilité : {rev}
-- Urgence : {urg}
-
-Résultat du moteur de qualification :
-- Type : {base.result.decision_type}
-- Processus recommandé : {base.result.process}
-- Modalités : {', '.join(base.result.recommended_modalities) or 'aucune'}
-- Gravure on-chain : {'recommandée' if base.result.recommend_onchain else 'non nécessaire'}
-
-Rédige une explication courte (2-3 phrases) qui explique pourquoi ce processus est recommandé \
-pour cette décision spécifique. Sois direct et concis. Réponds en français."""
-
-    try:
-        message = await client.messages.create(
-            model=settings.ANTHROPIC_MODEL,
-            max_tokens=300,
-            messages=[{"role": "user", "content": prompt}],
-        )
-        return message.content[0].text.strip()
-    except Exception as exc:
-        logger.warning("Claude explanation generation failed: %s — using fallback", exc)
-        return None
-
-
-# ---------------------------------------------------------------------------
-# Rule-based helpers (shared by sync and async paths)
+# Helpers
 # ---------------------------------------------------------------------------
 
 
diff --git a/backend/requirements.txt b/backend/requirements.txt
index 1c50e8c..8401843 100644
--- a/backend/requirements.txt
+++ b/backend/requirements.txt
@@ -12,7 +12,6 @@ substrate-interface==1.7.10
 py-sr25519-bindings==0.2.1
 base58==2.1.1
 httpx==0.28.1
-anthropic>=0.97.0
 aioipfs==0.7.1
 pytest==8.3.4
 pytest-asyncio==0.24.0