🛡️ 内容安全

class ContentSafetyFilter:
    def __init__(self):
        # 多级过滤器
        self.filters = [
            KeywordFilter(),           # 关键词匹配
            ClassifierFilter(),        # ML 分类器
            LLMJudgeFilter(),          # LLM 评估
            RuleBasedFilter(),         # 规则引擎
        ]
    
    def evaluate(self, text: str, context: dict) -> SafetyResult:
        """多层安全评估"""
        
        # Level 1: 快速规则检查
        if self.filters[0].detect(text):
            return SafetyResult(blocked=True, reason="keyword")
        
        # Level 2: ML 分类
        risk_score = self.filters[1].predict(text)
        if risk_score > 0.8:
            return SafetyResult(blocked=True, reason="ml_high_risk")
        
        # Level 3: LLM 深度评估
        llm_verdict = self.filters[2].judge(text, context)
        
        # Level 4: 规则检查
        if self.filters[3].check(text, context):
            return SafetyResult(blocked=True, reason="rule")
        
        return SafetyResult(blocked=False, score=risk_score)

class BiasEvaluator:
    def evaluate(self, dataset: Dataset) -> BiasReport:
        """全面偏见评估"""
        
        reports = {}
        
        # 1. 性别偏见
        gender_report = self.evaluate_gender_bias(dataset)
        
        # 2. 种族偏见
        racial_report = self.evaluate_racial_bias(dataset)
        
        # 3. 职业偏见
        occupation_report = self.evaluate_occupation_bias(dataset)
        
        # 4. 文化偏见
        cultural_report = self.evaluate_cultural_bias(dataset)
        
        # 综合评分
        overall_score = self.compute_overall_score(
            gender_report, racial_report,
            occupation_report, cultural_report
        )
        
        return BiasReport(
            gender=gender_report,
            racial=racial_report,
            occupation=occupation_report,
            cultural=cultural_report,
            overall=overall_score
        )

class FactChecker:
    def __init__(self, retriever, verifier):
        self.retriever = retriever  # RAG 检索器
        self.verifier = verifier    # LLM 验证器
    
    def check(self, text: str) -> FactCheckReport:
        claims = self.extract_claims(text)
        results = []
        
        for claim in claims:
            # 检索相关证据
            evidence = self.retriever.search(claim)
            
            # LLM 验证
            verdict = self.verifier.verify(claim, evidence)
            
            results.append(FactClaim(
                claim=claim,
                verdict=verdict,
                evidence=evidence[:3],
                confidence=verdict.confidence
            ))
        
        return FactCheckReport(
            text=text,
            claims=results,
            overall_reliability=self.compute_reliability(results)
        )