diff --git a/backend/planning_exercise_path_builder.py b/backend/planning_exercise_path_builder.py
index 7722146..752f1b9 100644
--- a/backend/planning_exercise_path_builder.py
+++ b/backend/planning_exercise_path_builder.py
@@ -110,7 +110,7 @@ class ProgressionPathSuggestRequest(BaseModel):
     include_path_qa: bool = True
     auto_rematch_after_qa: bool = True
     auto_refine_stage_spec: bool = True
-    max_rematch_rounds: int = Field(default=2, ge=0, le=3)
+    max_rematch_rounds: int = Field(default=3, ge=0, le=4)
     include_llm_path_qa: bool = True
     include_path_reorder: bool = True
     include_ai_gap_fill: bool = True
@@ -1297,6 +1297,23 @@ def _run_roadmap_rematch_loop(
     current_stripped = list(stripped_off_topic or [])
     use_initial_off_topic = not current_stripped
     off_topic_steps: List[Dict[str, Any]] = []
+    rejected_by_major: Dict[int, Set[int]] = {}
+
+    def _track_rejected(items: Sequence[Mapping[str, Any]]) -> None:
+        for item in items or []:
+            if not isinstance(item, dict):
+                continue
+            eid = item.get("exercise_id")
+            midx = item.get("roadmap_major_step_index")
+            if eid is None or midx is None:
+                continue
+            try:
+                rejected_by_major.setdefault(int(midx), set()).add(int(eid))
+            except (TypeError, ValueError):
+                continue
+
+    _track_rejected(off_topic_before_strip)
+    _track_rejected(current_stripped)
 
     for round_idx in range(max_rounds):
         mini_qa = run_multistage_path_qa(
@@ -1357,12 +1374,20 @@ def _run_roadmap_rematch_loop(
             slot_indices=slot_indices,
             rematch_reasons=rematch_reasons,
             match_slot_fn=_match_roadmap_slot,
+            rejected_by_major=rejected_by_major,
         )
         rematch_rounds += 1
         for entry in round_log:
             tagged = dict(entry)
             tagged["round"] = rematch_rounds
             rematch_log.append(tagged)
+            rid = entry.get("replaced_exercise_id")
+            midx = entry.get("roadmap_major_step_index")
+            if rid is not None and midx is not None:
+                try:
+                    rejected_by_major.setdefault(int(midx), set()).add(int(rid))
+                except (TypeError, ValueError):
+                    pass
 
         current_stripped = prune_stripped_after_rematch(current_stripped, round_log)
         roadmap_unfilled = _merge_rematch_unfilled(roadmap_unfilled, rematch_new_unfilled)
@@ -1374,6 +1399,7 @@ def _run_roadmap_rematch_loop(
             brief=semantic_brief,
             goal_query=goal_query,
         )
+        _track_rejected(off_topic_steps)
         if round_idx + 1 >= max_rounds:
             break
         if not off_topic_steps and not roadmap_unfilled:
diff --git a/backend/planning_exercise_path_qa.py b/backend/planning_exercise_path_qa.py
index d3e50a7..79f6b16 100644
--- a/backend/planning_exercise_path_qa.py
+++ b/backend/planning_exercise_path_qa.py
@@ -21,12 +21,15 @@ from planning_exercise_semantics import (
     _blob_from_fields,
     _blob_matches_stage_excludes,
     brief_to_summary_dict,
+    build_stage_match_brief,
     exercise_passes_path_semantic_gate,
     exercise_passes_stage_learning_goal_gate,
     exercise_passes_technique_path_scope,
+    merge_stage_exclude_phrases,
     resolve_path_anti_patterns,
     resolve_path_primary_topic,
     score_exercise_semantic_relevance,
+    score_exercise_stage_fit,
     semantic_brief_for_stage,
     step_phase_for_index,
     technique_sibling_excludes,
@@ -442,8 +445,13 @@ def detect_off_topic_steps(
             bundle["goal"],
             bundle["variant_names"],
         )
-        step_anti = list(step.get("roadmap_anti_patterns") or []) + path_anti
-        if step_anti and _blob_matches_stage_excludes(blob, step_anti):
+        step_anti_raw = list(step.get("roadmap_anti_patterns") or [])
+        stage_goal_pre = (step.get("roadmap_learning_goal") or "").strip()
+        exclude_phrases = merge_stage_exclude_phrases(
+            stage_goal_pre,
+            [*step_anti_raw, *path_anti],
+        )
+        if exclude_phrases and _blob_matches_stage_excludes(blob, exclude_phrases):
             off_topic.append(
                 _with_roadmap_major_index(
                     step,
@@ -459,7 +467,6 @@ def detect_off_topic_steps(
                 )
             )
             continue
-        stage_goal_pre = (step.get("roadmap_learning_goal") or "").strip()
         primary = (
             resolve_path_primary_topic(
                 goal_query or "",
@@ -512,6 +519,26 @@ def detect_off_topic_steps(
             step_phase=phase,
         )
         stage_anti = list(step.get("roadmap_anti_patterns") or [])
+        stage_match_brief = (
+            build_stage_match_brief(
+                learning_goal=stage_goal,
+                anti_patterns=stage_anti or None,
+                phase=phase or None,
+            )
+            if stage_goal
+            else None
+        )
+        stage_sem = 0.0
+        stage_reasons: List[str] = []
+        if stage_match_brief:
+            stage_sem, stage_reasons = score_exercise_stage_fit(
+                title=bundle["title"],
+                summary=bundle["summary"],
+                goal=bundle["goal"],
+                variant_names=bundle["variant_names"],
+                stage_brief=stage_match_brief,
+                step_phase=phase,
+            )
         if stage_goal and not exercise_passes_stage_learning_goal_gate(
             learning_goal=stage_goal,
             title=bundle["title"],
@@ -520,6 +547,15 @@ def detect_off_topic_steps(
             semantic_score=sem,
             anti_patterns=stage_anti or None,
         ):
+            reasons = [
+                r
+                for r in stage_reasons
+                if r and r != "Kern-Thema der Anfrage im Übungstext"
+            ]
+            if not reasons:
+                reasons = [
+                    f"Stufen-Fit zu schwach ({stage_sem:.2f}) für „{stage_goal[:80]}“"
+                ]
             off_topic.append(
                 _with_roadmap_major_index(
                     step,
@@ -527,11 +563,11 @@ def detect_off_topic_steps(
                         "step_index": idx,
                         "exercise_id": int(step["exercise_id"]),
                         "title": step.get("title") or bundle["title"],
-                        "semantic_score": round(sem, 4),
+                        "semantic_score": round(stage_sem, 4),
                         "expected_phase": phase,
                         "issue": "stage_mismatch",
                         "roadmap_learning_goal": stage_goal,
-                        "reasons": sem_reasons[:3],
+                        "reasons": reasons[:3],
                     },
                 )
             )
diff --git a/backend/planning_exercise_semantics.py b/backend/planning_exercise_semantics.py
index 501c884..c140075 100644
--- a/backend/planning_exercise_semantics.py
+++ b/backend/planning_exercise_semantics.py
@@ -813,6 +813,70 @@ def _expand_stage_exclude_phrase(phrase: str) -> List[str]:
     return out[:12]
 
 
+def is_trainer_stage_anti_marker(raw: str) -> bool:
+    """Trainer-/QS-Marker — nicht als Negationsphrase parsen."""
+    norm = _normalize_phrase(str(raw or ""))
+    if not norm:
+        return False
+    stripped = re.sub(r"[„“\"'«»]", "", norm)
+    stripped = re.sub(r"\s+", " ", stripped).strip()
+    if stripped.startswith("keine übung wie") or stripped.startswith("keine uebung wie"):
+        return True
+    return stripped.startswith("qs-hinweis")
+
+
+def merge_stage_exclude_phrases(
+    learning_goal: str,
+    anti_patterns: Optional[Sequence[str]] = None,
+) -> List[str]:
+    """
+    Ausschlussphrasen für Stufen-Gates — Negationen nur aus dem Lernziel expandieren,
+    explizite anti_patterns unverändert (ohne Trainer-Marker erneut zu parsen).
+    """
+    lg = (learning_goal or "").strip()
+    exclude: List[str] = []
+    if len(lg) >= 3:
+        for item in parse_stage_goal_constraints(lg).exclude_phrases:
+            if item and item not in exclude:
+                exclude.append(item)
+    markers: List[str] = []
+    for raw in anti_patterns or []:
+        s = str(raw or "").strip()
+        if not s:
+            continue
+        if is_trainer_stage_anti_marker(s):
+            if s not in markers:
+                markers.append(s[:200])
+            continue
+        norm = _normalize_phrase(s)
+        if norm and norm not in exclude:
+            exclude.append(norm)
+    for marker in markers:
+        if marker not in exclude:
+            exclude.append(marker)
+    return exclude[:16]
+
+
+def stage_focus_phrases_from_learning_goal(learning_goal: str) -> List[str]:
+    """Mehrwort-Schwerpunkte aus Stufen-Lernziel für Fit-Scoring."""
+    lg = (learning_goal or "").strip()
+    if len(lg) < 3:
+        return []
+    tokens = _significant_stage_tokens(lg, strip_negated=True)
+    phrases: List[str] = []
+    for tok in tokens:
+        if len(tok) >= 5 and tok not in phrases:
+            phrases.append(tok)
+    for i in range(len(tokens) - 1):
+        pair = f"{tokens[i]} {tokens[i + 1]}"
+        if len(pair) >= 8 and pair not in phrases:
+            phrases.append(pair)
+    norm_lg = _normalize_phrase(lg)
+    if len(norm_lg) >= 8 and norm_lg not in phrases:
+        phrases.insert(0, norm_lg[:120])
+    return phrases[:8]
+
+
 def _significant_stage_tokens(learning_goal: str, *, strip_negated: bool = True) -> List[str]:
     """Wörter aus Stufen-Lernziel für Text-Match (ohne Füllwörter, ohne Negationssegmente)."""
     text = _normalize_phrase(learning_goal)
@@ -850,9 +914,11 @@ def parse_stage_goal_constraints(
                 exclude.extend(_expand_stage_exclude_phrase(chunk))
 
     for raw in anti_patterns or []:
+        if is_trainer_stage_anti_marker(str(raw or "")):
+            continue
         s = _normalize_phrase(str(raw or ""))
-        if s:
-            exclude.extend(_expand_stage_exclude_phrase(s))
+        if s and s not in exclude:
+            exclude.append(s)
 
     positive = _significant_stage_tokens(lg, strip_negated=True)
     focus_hits = [t for t in positive if t in _STAGE_FOCUS_TOKENS]
@@ -997,7 +1063,7 @@ def build_stage_match_brief(
         for expanded in _expand_stage_exclude_phrase(str(raw or "")):
             if expanded and expanded not in merged_anti:
                 merged_anti.append(expanded)
-    constraints = parse_stage_goal_constraints(lg, merged_anti)
+    constraints = parse_stage_goal_constraints(lg)
     must: List[str] = []
     norm_lg = _normalize_phrase(lg)
     if primary_path and primary_path not in must:
@@ -1031,11 +1097,13 @@ def build_stage_match_brief(
     if ph:
         arc.append(ph)
 
+    exclude_phrases = merge_stage_exclude_phrases(lg, merged_anti)
+
     return PlanningSemanticBrief(
         primary_topic="",
         topic_type="focus",
         must_phrases=must[:12],
-        exclude_phrases=list(constraints.exclude_phrases)[:12],
+        exclude_phrases=exclude_phrases[:12],
         development_arc=arc[:4],
         retrieval_query=" ".join(p for p in retrieval_parts if p)[:500],
         semantic_strength=0.78,
@@ -1062,19 +1130,36 @@ def score_exercise_stage_fit(
         step_phase=step_phase,
     )
     blob = _blob_from_fields(title, summary, goal, variant_names or [])
-    focus_tokens = [
-        t
-        for t in (stage_brief.must_phrases or [])
-        if t and " " not in t and len(t) >= 4
-    ][:6]
-    if focus_tokens:
-        hits = sum(1 for t in focus_tokens if _phrase_in_blob(t, blob))
-        ratio = hits / len(focus_tokens)
-        bonus = 0.28 * ratio
+    lg_hint = ""
+    for part in (stage_brief.retrieval_query or "").split("|"):
+        part = part.strip()
+        if part.lower().startswith("lernziel:"):
+            lg_hint = part.split(":", 1)[-1].strip()
+            break
+    if not lg_hint:
+        for mp in stage_brief.must_phrases or []:
+            if mp and len(_normalize_phrase(mp)) >= 8:
+                lg_hint = mp
+                break
+    focus_phrases = stage_focus_phrases_from_learning_goal(lg_hint) if lg_hint else []
+    if not focus_phrases:
+        focus_phrases = [
+            t
+            for t in (stage_brief.must_phrases or [])
+            if t and len(_normalize_phrase(t)) >= 5
+        ][:6]
+    if focus_phrases:
+        hits = sum(1 for p in focus_phrases if _phrase_in_blob(p, blob))
+        ratio = hits / len(focus_phrases)
+        bonus = 0.32 * ratio
         if bonus > 0:
             score = min(1.0, score + bonus)
-            if hits >= max(1, len(focus_tokens) // 2):
+            if hits >= max(1, len(focus_phrases) // 2):
                 reasons = ["Stufen-Schwerpunkte im Übungstext", *reasons]
+    learning_goal_for_equiv = lg_hint or (stage_brief.must_phrases[0] if stage_brief.must_phrases else "")
+    if learning_goal_for_equiv and exercise_title_equivalent_to_stage_goal(title, learning_goal_for_equiv):
+        score = max(score, 0.42)
+        reasons = ["Titel entspricht Stufen-Lernziel", *reasons]
     return max(0.0, min(1.0, round(score, 4))), reasons[:4]
 
 
@@ -1099,11 +1184,13 @@ def exercise_passes_stage_fit(
         return True
 
     blob = _blob_from_fields(title, summary, goal, [])
-    constraints = parse_stage_goal_constraints(lg, anti_patterns)
-    if constraints.exclude_phrases and _blob_matches_stage_excludes(blob, constraints.exclude_phrases):
+    exclude_phrases = merge_stage_exclude_phrases(lg, anti_patterns)
+    if exclude_phrases and _blob_matches_stage_excludes(blob, exclude_phrases):
         return False
 
     title_equiv = exercise_title_equivalent_to_stage_goal(title, learning_goal or lg)
+    if title_equiv:
+        return True
 
     primary_path = (path_primary_topic or "").strip()
     if not primary_path and lg:
@@ -1293,10 +1380,8 @@ def _pick_roadmap_rank_fallback(
         summary = str(hit.get("summary") or "")
         goal_text = str(hit.get("goal") or hit.get("exercise_goal") or "")
         blob = _blob_from_fields(title, summary, goal_text, [])
-        constraints = parse_stage_goal_constraints(stage_goal, stage_anti_patterns)
-        if constraints.exclude_phrases and _blob_matches_stage_excludes(
-            blob, constraints.exclude_phrases
-        ):
+        exclude_phrases = merge_stage_exclude_phrases(stage_goal, stage_anti_patterns)
+        if exclude_phrases and _blob_matches_stage_excludes(blob, exclude_phrases):
             continue
         title_equiv = exercise_title_equivalent_to_stage_goal(title, stage_goal)
         primary = (path_primary_topic or "").strip()
@@ -1465,8 +1550,11 @@ __all__ = [
     "resolve_path_primary_topic",
     "resolve_path_anti_patterns",
     "exercise_passes_stage_learning_goal_gate",
+    "is_trainer_stage_anti_marker",
     "merge_semantic_brief_llm",
+    "merge_stage_exclude_phrases",
     "parse_stage_goal_constraints",
+    "stage_focus_phrases_from_learning_goal",
     "pick_best_path_hit",
     "exercise_passes_technique_path_scope",
     "score_exercise_stage_fit",
diff --git a/backend/planning_path_refine_stage.py b/backend/planning_path_refine_stage.py
index 04ee7dd..0ee9987 100644
--- a/backend/planning_path_refine_stage.py
+++ b/backend/planning_path_refine_stage.py
@@ -2,17 +2,17 @@
 Phase C: Stufen-Spec verfeinern nach stage_mismatch, dann Rematch.
 
 Deterministisch — keine LLM-Ratelosigkeit. Schärft anti_patterns / success_criteria
-aus QS-Finding, schließt abgelehnte Übung aus, übernimmt Pfad-Ausschlüsse.
+aus QS-Finding, schließt abgelehnte Übung aus.
 """
 from __future__ import annotations
 
 from typing import Any, Dict, List, Mapping, Optional, Sequence, Set, Tuple
 
 from planning_exercise_semantics import (
-    PlanningSemanticBrief,
-    build_stage_match_brief,
+    is_trainer_stage_anti_marker,
+    merge_stage_exclude_phrases,
     parse_stage_goal_constraints,
-    resolve_path_anti_patterns,
+    stage_focus_phrases_from_learning_goal,
 )
 from planning_progression_roadmap import ProgressionRoadmapContext, StageSpecArtifact
 
@@ -79,65 +79,57 @@ def _append_unique_strings(dest: List[str], items: Sequence[str], *, limit: int
     return out
 
 
+def _rejected_exercise_marker(title: str) -> str:
+    return f"keine Übung wie „{title[:120]}“"
+
+
 def refine_stage_spec_artifact(
     spec: StageSpecArtifact,
     *,
     finding: Mapping[str, Any],
-    goal_query: str,
-    semantic_brief: Optional[PlanningSemanticBrief] = None,
+    goal_query: str = "",
+    semantic_brief: Optional[Any] = None,
     path_anti_patterns: Optional[Sequence[str]] = None,
 ) -> Tuple[StageSpecArtifact, List[str]]:
     """
     Schärft eine StageSpec aus QS-Finding. Returns (neue Spec, Änderungsliste).
+
+    Pfad-Ausschlüsse werden beim Match separat gemerged — nicht in stage_spec duplizieren.
     """
+    del goal_query, semantic_brief, path_anti_patterns
     learning_goal = (
         str(finding.get("roadmap_learning_goal") or spec.learning_goal or "").strip()
         or spec.learning_goal
     )
-    anti = list(spec.anti_patterns or [])
+    anti = [a for a in list(spec.anti_patterns or []) if not is_trainer_stage_anti_marker(a)]
     success = list(spec.success_criteria or [])
     changes: List[str] = []
 
     rejected_title = str(finding.get("title") or "").strip()
     if rejected_title:
-        marker = f"keine Übung wie „{rejected_title[:120]}“"
+        marker = _rejected_exercise_marker(rejected_title)
         if marker not in anti:
             anti.append(marker)
             changes.append(f"Ausschluss abgelehnter Übung: {rejected_title[:80]}")
 
-    path_anti = list(path_anti_patterns or [])
-    if not path_anti and semantic_brief is not None:
-        path_anti = resolve_path_anti_patterns(goal_query, semantic_brief=semantic_brief)
-    merged_anti = _append_unique_strings(anti, path_anti)
-    if len(merged_anti) > len(anti):
-        changes.append("Pfad-Ausschlüsse in Stufen-anti_patterns übernommen")
-    anti = merged_anti
-
-    constraints = parse_stage_goal_constraints(learning_goal, anti)
-    for phrase in constraints.exclude_phrases or []:
+    goal_excludes = parse_stage_goal_constraints(learning_goal).exclude_phrases
+    for phrase in goal_excludes or []:
         if phrase and phrase not in anti:
             anti.append(phrase)
             changes.append(f"Ausschluss aus Lernziel: {phrase[:60]}")
 
-    stage_brief = build_stage_match_brief(
-        learning_goal=learning_goal,
-        anti_patterns=anti,
-        success_criteria=list(spec.success_criteria or []),
-        load_profile=list(spec.load_profile or []),
-    )
-    for phrase in (stage_brief.must_phrases or [])[:4]:
-        p = str(phrase or "").strip()
-        if len(p) < 4:
-            continue
-        crit = f"Bezug zu Stufen-Lernziel: {p[:100]}"
+    for phrase in stage_focus_phrases_from_learning_goal(learning_goal):
+        crit = f"Bezug zu Stufen-Lernziel: {phrase[:100]}"
         if crit not in success:
             success.append(crit)
-            changes.append(f"Erfolgskriterium: {p[:60]}")
+            changes.append(f"Erfolgskriterium: {phrase[:60]}")
 
     for raw in finding.get("reasons") or []:
         r = str(raw or "").strip()
         if len(r) < 8:
             continue
+        if r == "Kern-Thema der Anfrage im Übungstext":
+            continue
         crit = f"QS-Hinweis: {r[:120]}"
         if crit not in success:
             success.append(crit)
@@ -157,7 +149,7 @@ def refine_stage_spec_artifact(
         load_profile=list(spec.load_profile or []),
         exercise_type=spec.exercise_type,
         success_criteria=success[:8],
-        anti_patterns=anti[:14],
+        anti_patterns=merge_stage_exclude_phrases(learning_goal, anti)[:14],
     )
     return refined, changes
 
@@ -168,13 +160,14 @@ def apply_stage_spec_refinements(
     optimization_hints: Sequence[Mapping[str, Any]],
     off_topic_steps: Sequence[Mapping[str, Any]],
     goal_query: str,
-    semantic_brief: Optional[PlanningSemanticBrief] = None,
+    semantic_brief: Optional[Any] = None,
 ) -> Tuple[List[StageSpecArtifact], List[Dict[str, Any]]]:
     """
     Wendet refine_stage_spec auf betroffene Slots an (mutiert stage_specs in ctx).
 
     Returns: (stage_specs, refine_log)
     """
+    del goal_query, semantic_brief
     stage_specs = list(roadmap_ctx.stage_specs or [])
     if not stage_specs:
         return stage_specs, []
@@ -187,10 +180,8 @@ def apply_stage_spec_refinements(
     if not targets:
         return stage_specs, []
 
-    path_anti = resolve_path_anti_patterns(goal_query, semantic_brief=semantic_brief)
     spec_by_major = {int(s.major_step_index): s for s in stage_specs}
     refine_log: List[Dict[str, Any]] = []
-    refined_majors: Set[int] = set()
 
     for midx in sorted(targets):
         spec = spec_by_major.get(int(midx))
@@ -199,20 +190,18 @@ def apply_stage_spec_refinements(
         refined_spec, changes = refine_stage_spec_artifact(
             spec,
             finding=targets[midx],
-            goal_query=goal_query,
-            semantic_brief=semantic_brief,
-            path_anti_patterns=path_anti,
         )
         if not changes:
             continue
         spec_by_major[int(midx)] = refined_spec
-        refined_majors.add(int(midx))
+        rejected_id = targets[midx].get("exercise_id")
         refine_log.append(
             {
                 "roadmap_major_step_index": int(midx),
                 "action": "refined",
                 "issue": "stage_mismatch",
                 "rejected_title": targets[midx].get("title"),
+                "rejected_exercise_id": int(rejected_id) if rejected_id else None,
                 "changes": changes[:6],
                 "reason": (changes[0] if changes else "refine_stage_spec")[:400],
             }
diff --git a/backend/planning_path_rematch.py b/backend/planning_path_rematch.py
index 4bc19dd..9adcb5c 100644
--- a/backend/planning_path_rematch.py
+++ b/backend/planning_path_rematch.py
@@ -115,6 +115,7 @@ def rematch_roadmap_slots(
     slot_indices: Set[int],
     rematch_reasons: Mapping[int, str],
     match_slot_fn,
+    rejected_by_major: Optional[Mapping[int, Set[int]]] = None,
 ) -> Tuple[List[Dict[str, Any]], List[Dict[str, Any]], List[Tuple[int, StageSpecArtifact]]]:
     """
     Ersetzt nur betroffene Slots; andere Schritte und used-Set bleiben konsistent.
@@ -152,6 +153,9 @@ def rematch_roadmap_slots(
         }
         if old and old.get("exercise_id") is not None:
             used.add(int(old["exercise_id"]))
+        for rejected_id in rejected_by_major.get(int(major_idx), set()) if rejected_by_major else set():
+            if rejected_id > 0:
+                used.add(int(rejected_id))
         planned_ids, anchor_id, anchor_variant_id = _context_before_major(
             steps_by_major, int(major_idx)
         )
diff --git a/backend/tests/test_planning_path_refine_stage.py b/backend/tests/test_planning_path_refine_stage.py
index 529e278..5cbdaca 100644
--- a/backend/tests/test_planning_path_refine_stage.py
+++ b/backend/tests/test_planning_path_refine_stage.py
@@ -1,5 +1,4 @@
 """Tests Phase C — refine_stage_spec nach stage_mismatch."""
-from planning_exercise_semantics import build_semantic_brief
 from planning_path_refine_stage import (
     apply_stage_spec_refinements,
     collect_refine_stage_targets,
@@ -49,17 +48,14 @@ def test_refine_stage_spec_adds_rejected_title_and_criteria():
         "roadmap_learning_goal": spec.learning_goal,
         "reasons": ["Semantik zu schwach für Stufen-Lernziel"],
     }
-    brief = build_semantic_brief("Mawashi Geri Kumite")
     refined, changes = refine_stage_spec_artifact(
         spec,
         finding=finding,
-        goal_query="Mawashi Geri ohne Kumite",
-        semantic_brief=brief,
     )
     assert changes
-    assert any("Mawashi Trittpräzision" in a for a in refined.anti_patterns)
+    assert any("Mawashi" in a and "Tritt" in a for a in refined.anti_patterns)
     assert refined.success_criteria
-    assert refined.anti_patterns != spec.anti_patterns or refined.success_criteria != spec.success_criteria
+    assert not any("anderetechnikals" in a.replace(" ", "") for a in refined.anti_patterns)
 
 
 def test_apply_stage_spec_refinements_mutates_context():
@@ -81,7 +77,6 @@ def test_apply_stage_spec_refinements_mutates_context():
             }
         ],
         goal_query="Mawashi Geri",
-        semantic_brief=build_semantic_brief("Mawashi Geri"),
     )
     assert len(log) == 1
     assert log[0]["action"] == "refined"
diff --git a/backend/tests/test_planning_stage_anti_patterns.py b/backend/tests/test_planning_stage_anti_patterns.py
new file mode 100644
index 0000000..5c97fb9
--- /dev/null
+++ b/backend/tests/test_planning_stage_anti_patterns.py
@@ -0,0 +1,54 @@
+"""Tests für Stufen-Ausschlüsse und Anti-Pattern-Sanitizer."""
+from planning_exercise_semantics import (
+    exercise_passes_stage_fit,
+    is_trainer_stage_anti_marker,
+    merge_stage_exclude_phrases,
+    parse_stage_goal_constraints,
+    score_exercise_stage_fit,
+    build_stage_match_brief,
+)
+
+
+def test_trainer_anti_marker_not_reparsed_as_negation():
+    marker = 'keine Übung wie „One Leg Squat“'
+    assert is_trainer_stage_anti_marker(marker)
+    excludes = merge_stage_exclude_phrases(
+        "Gleichgewichtstritt Mae-Geri",
+        [marker, "kumite"],
+    )
+    assert "onelegsquat" not in "".join(excludes).replace(" ", "")
+    assert "kumite" in excludes
+
+
+def test_parse_stage_goal_constraints_skips_trainer_markers_in_anti():
+    marker = 'keine Übung wie „Erlernen des Mae Geri aus zusammengesetzten Einzelbewegungen“'
+    result = parse_stage_goal_constraints(
+        "Koordination ohne Kumite",
+        [marker],
+    )
+    joined = " ".join(result.exclude_phrases)
+    assert "keine uebung wie" not in joined
+    assert "kumite" in joined
+
+
+def test_title_equivalent_passes_stage_fit_despite_low_semantic():
+    goal = "Gleichgewichtstritt Mae-Geri"
+    assert exercise_passes_stage_fit(
+        learning_goal=goal,
+        title="Gleichgewichtstritt Mae-Geri",
+        summary="Balance und Treffpunkt variieren.",
+        goal="Mae Geri aus Stand.",
+        stage_semantic_score=0.05,
+    )
+
+
+def test_stage_focus_scoring_rewards_learning_goal_tokens():
+    goal = "Erlernen des Mae Geri aus zusammengesetzten Einzelbewegungen"
+    brief = build_stage_match_brief(learning_goal=goal)
+    score, reasons = score_exercise_stage_fit(
+        title="Mae Geri aus Einzelteilen",
+        summary="Zusammensetzung aus Schritt und Armschwingung.",
+        goal="Einzelbewegungen verbinden.",
+        stage_brief=brief,
+    )
+    assert score >= 0.25
diff --git a/backend/version.py b/backend/version.py
index 83b0284..90c12e1 100644
--- a/backend/version.py
+++ b/backend/version.py
@@ -1,6 +1,6 @@
 # Shinkan Jinkendo Version Information
 
-APP_VERSION = "0.8.227"
+APP_VERSION = "0.8.228"
 BUILD_DATE = "2026-05-22"
 DB_SCHEMA_VERSION = "20260607090"
 
@@ -38,7 +38,7 @@ MODULE_VERSIONS = {
     "skill_profiles": "1.0.0",  # Phase 3: gewichtetes Fähigkeiten-Profil + skill-discovery/suggestions
     "methods": "0.1.0",
     "exercises": "2.37.1",  # KI-Endpoints: feature_usage nach ai_calls consume
-    "planning_exercise_suggest": "0.23.2",  # Phase C: refine_stage_spec bei stage_mismatch vor Rematch
+    "planning_exercise_suggest": "0.23.3",  # Stufen-Match: saubere Anti-Patterns, Fit-Scoring, Rematch-Akkumulation
     "training_units": "0.4.0",  # POST .../publish-to-framework: Ablauf aus geplanter Einheit → Rahmen-Slot-Blueprint
     "training_programs": "0.1.0",
     "planning": "0.15.0",  # Vorlagen: Strukturvorschau, Bearbeiten inkl. Split-Sessions + Beschreibung
diff --git a/frontend/src/utils/progressionGraphDraft.js b/frontend/src/utils/progressionGraphDraft.js
index 81f7983..1a6489e 100644
--- a/frontend/src/utils/progressionGraphDraft.js
+++ b/frontend/src/utils/progressionGraphDraft.js
@@ -113,7 +113,10 @@ export function formatRefineLogEntry(entry) {
 }
 
 export function hasRematchSlotHints(pathQa) {
-  return (pathQa?.optimization_hints || []).some((h) => h?.action === 'rematch_slot')
+  return (pathQa?.optimization_hints || []).some((h) => {
+    const action = h?.action
+    return action === 'rematch_slot' || action === 'refine_stage_spec'
+  })
 }
 
 function createEmptySlot(index) {