diff --git a/decision_engine/src/decision_engine/core.py b/decision_engine/src/decision_engine/core.py
new file mode 100644
index 0000000..59208d4
--- /dev/null
+++ b/decision_engine/src/decision_engine/core.py
@@ -0,0 +1,113 @@
+from __future__ import annotations
+
+import logging
+from typing import Dict, Any, Literal
+
+logger = logging.getLogger(__name__)
+logger.setLevel(logging.INFO)
+
+
+class EvaluationError(Exception):
+    """Custom exception raised when evaluation cannot be performed due to invalid input."""
+
+
+class InputValidationError(EvaluationError):
+    pass
+
+
+Decision = Literal["PASS", "WARN", "FAIL"]
+
+
+def _validate_run_data(run_data: Dict[str, Any]) -> None:
+    required_fields = [
+        "warn_rate",
+        "unknown_rate",
+        "pinned",
+        "unknown_class",
+    ]
+
+    for field in required_fields:
+        if field not in run_data:
+            raise InputValidationError(f"Missing required field: {field}")
+
+    if not isinstance(run_data["warn_rate"], (int, float)):
+        raise InputValidationError("warn_rate must be a numeric value.")
+    if not isinstance(run_data["unknown_rate"], (int, float)):
+        raise InputValidationError("unknown_rate must be a numeric value.")
+    if not isinstance(run_data["pinned"], bool):
+        raise InputValidationError("pinned must be a boolean value.")
+    if not isinstance(run_data["unknown_class"], str):
+        raise InputValidationError("unknown_class must be a string.")
+
+
+# Placeholder thresholds for logical example; actual thresholds are read externally in CLI.
+_default_thresholds = {
+    "warn_rate": {"warn": 5.0, "fail": 10.0},
+    "unknown_rate": {"warn": 2.0, "fail": 5.0},
+}
+
+
+def evaluate_run(run_data: Dict[str, Any]) -> Dict[str, str]:
+    """Evaluate a single run's metrics and decide PASS/WARN/FAIL.
+
+    Args:
+        run_data: dict - contains fields warn_rate, unknown_rate, pinned, unknown_class, prev_label(optional)
+
+    Returns:
+        dict: {"final_decision": str, "reason": str}
+
+    Raises:
+        InputValidationError: if run_data fields are missing or of invalid types.
+    """
+    _validate_run_data(run_data)
+
+    warn_rate = float(run_data["warn_rate"])
+    unknown_rate = float(run_data["unknown_rate"])
+    pinned = run_data["pinned"]
+    unknown_class = run_data["unknown_class"]
+    prev_label = run_data.get("prev_label")
+
+    try:
+        w_warn = _default_thresholds["warn_rate"]["warn"]
+        w_fail = _default_thresholds["warn_rate"]["fail"]
+        u_warn = _default_thresholds["unknown_rate"]["warn"]
+        u_fail = _default_thresholds["unknown_rate"]["fail"]
+    except KeyError as e:
+        raise EvaluationError(f"Threshold configuration missing key: {e}") from e
+
+    # Decision logic
+    decision: Decision
+    reason: str
+
+    if warn_rate >= w_fail or unknown_rate >= u_fail:
+        decision = "FAIL"
+        reason = f"Run exceeded fail thresholds (warn_rate={warn_rate}, unknown_rate={unknown_rate})."
+    elif warn_rate >= w_warn or unknown_rate >= u_warn:
+        decision = "WARN"
+        reason = f"Run exceeded warning thresholds (warn_rate={warn_rate}, unknown_rate={unknown_rate})."
+    else:
+        decision = "PASS"
+        reason = "Run metrics are within acceptable thresholds."
+
+    # Adjustments for pinned or unknown_class abnormalities
+    if pinned and decision == "FAIL":
+        decision = "WARN"
+        reason += " Pinned run downgraded from FAIL to WARN."
+    if "violation" in unknown_class.lower() and decision == "PASS":
+        decision = "WARN"
+        reason += " Unknown class contains violation, upgraded to WARN."
+
+    # Integrate previous label as context info (no override logic by default)
+    if prev_label:
+        reason += f" Previous label: {prev_label}."
+
+    logger.debug(
+        "Evaluated run → decision=%s, reason=%s | data=%s",
+        decision,
+        reason,
+        run_data,
+    )
+
+    assert decision in {"PASS", "WARN", "FAIL"}, "Invalid decision value computed."
+
+    return {"final_decision": decision, "reason": reason}