Veris

"""
Veris: Objective AI Evaluator & LLM Regression CI Gate
"""
import statistics
from typing import List, Dict, Any
from pydantic import BaseModel, Field

class CandidateOutput(BaseModel):
    evaluator_id: str
    faithfulness_score: float
    relevance_score: float
    hallucination_flag: bool

class EvalGateReport(BaseModel):
    ci_passed: bool
    consensus_score: float
    variance: float
    hallucination_rate: float

class VerisEvaluator:
    def __init__(self, min_consensus: float = 0.85, max_variance: float = 0.05):
        self.min_consensus = min_consensus
        self.max_variance = max_variance

    def compute_jury_consensus(self, scores: List[CandidateOutput]) -> EvalGateReport:
        composite_scores = [(s.faithfulness_score + s.relevance_score) / 2.0 for s in scores]
        mean_score = statistics.mean(composite_scores)
        variance = statistics.variance(composite_scores) if len(composite_scores) > 1 else 0.0
        hallucination_rate = sum(1 for s in scores if s.hallucination_flag) / len(scores)

        passed = (mean_score >= self.min_consensus) and (variance <= self.max_variance) and (hallucination_rate == 0.0)

        return EvalGateReport(
            ci_passed=passed,
            consensus_score=round(mean_score, 3),
            variance=round(variance, 4),
            hallucination_rate=round(hallucination_rate, 2)
        )