"""Offline, dependency-free demonstration of defensive LLM response checks.""" from dataclasses import dataclass, asdict import re @dataclass(frozen=True) class Evaluation: passed: bool flags: tuple[str, ...] response_chars: int def evaluate_response(response: str) -> Evaluation: """Apply illustrative checks; not a comprehensive safety classifier.""" if not isinstance(response, str): raise TypeError('response must be a string') flags = [] if not response.strip(): flags.append('empty_response') if re.search(r'(?i)\b(?:system prompt|ignore previous instructions)\b', response): flags.append('review_instruction_reference') return Evaluation(passed=not flags, flags=tuple(flags), response_chars=len(response)) def evaluate_batch(responses: list[str]) -> list[dict]: return [asdict(evaluate_response(item)) for item in responses]