diff --git a/frozen_runs_analysis/src/frozen_runs_analysis/core.py b/frozen_runs_analysis/src/frozen_runs_analysis/core.py
new file mode 100644
index 0000000..543b74b
--- /dev/null
+++ b/frozen_runs_analysis/src/frozen_runs_analysis/core.py
@@ -0,0 +1,96 @@
+from __future__ import annotations
+
+import logging
+from dataclasses import dataclass, asdict
+from typing import Dict, Any, List
+import pandas as pd
+from statistics import mean, pstdev, quantiles
+
+
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.INFO)
+
+
+class RunDataValidationError(Exception):
+    """Raised when RunData validation fails."""
+    pass
+
+
+@dataclass
+class RunData:
+    run_id: str
+    status: str
+    sanity_checks: Dict[str, Any]
+    config_hash: str
+
+    @classmethod
+    def validate(cls, data: Dict[str, Any]) -> 'RunData':
+        required_fields = {"run_id", "status", "sanity_checks", "config_hash"}
+        missing = required_fields - data.keys()
+        if missing:
+            raise RunDataValidationError(f"Missing fields in RunData: {', '.join(missing)}")
+        if not isinstance(data.get("sanity_checks"), dict):
+            raise RunDataValidationError("Field 'sanity_checks' must be a dict.")
+        if not isinstance(data.get("run_id"), str):
+            raise RunDataValidationError("Field 'run_id' must be a str.")
+        if not isinstance(data.get("status"), str):
+            raise RunDataValidationError("Field 'status' must be a str.")
+        if not isinstance(data.get("config_hash"), str):
+            raise RunDataValidationError("Field 'config_hash' must be a str.")
+        return cls(
+            run_id=data["run_id"],
+            status=data["status"],
+            sanity_checks=data["sanity_checks"],
+            config_hash=data["config_hash"]
+        )
+
+
+def analyse_frozen_runs(data: List[RunData]) -> Dict[str, Any]:
+    """Analysiert Frozen-Run-Daten, berechnet Kennzahlen zur Stabilität und Häufigkeiten."""
+    # Validierung
+    if not data:
+        logger.warning("Keine Daten zur Analyse übergeben.")
+        return {}
+
+    valid_runs: List[RunData] = []
+    for item in data:
+        if isinstance(item, dict):
+            try:
+                valid_runs.append(RunData.validate(item))
+            except RunDataValidationError as e:
+                logger.error(f"Ungültiger Dateneintrag ignoriert: {e}")
+        elif isinstance(item, RunData):
+            valid_runs.append(item)
+        else:
+            logger.error(f"Unbekannter Datentyp in Analyse: {type(item).__name__}")
+
+    if not valid_runs:
+        logger.warning("Keine validen Läufe nach Validierung.")
+        return {}
+
+    df = pd.DataFrame([asdict(r) for r in valid_runs])
+
+    # Einfacher Sanity-Score: Anzahl Fehler in sanity_checks
+    df['sanity_error_count'] = df['sanity_checks'].apply(lambda d: sum(1 for v in d.values() if v not in (None, True, False) and not v))
+
+    # Flip-Flop-Frequenz per config_hash und status
+    hash_status = df.groupby('config_hash')['status'].nunique()
+    flip_flop_rate = (hash_status > 1).mean()
+
+    # Statistische Verteilungen
+    sanity_counts = df['sanity_error_count'].tolist()
+    avg_errors = mean(sanity_counts)
+    std_errors = pstdev(sanity_counts) if len(sanity_counts) > 1 else 0.0
+    p95 = quantiles(sanity_counts, n=100)[94] if len(sanity_counts) >= 20 else max(sanity_counts)
+
+    result = {
+        "total_runs": len(df),
+        "unique_hashes": df['config_hash'].nunique(),
+        "flip_flop_rate": round(float(flip_flop_rate), 4),
+        "avg_sanity_errors": round(avg_errors, 4),
+        "std_sanity_errors": round(std_errors, 4),
+        "p95_sanity_errors": round(float(p95), 4),
+    }
+
+    logger.info(f"Analyse abgeschlossen: {result}")
+    return result