core/aggregator.py

# core/aggregator.py
# Combines BERT score + stylometric score into a final probability.

from core.stylometrics import compute_stylometric_score
from core.explainer import find_ai_phrases, generate_explanation
from config.settings import settings


def aggregate_scores(
    bert_result: dict,
    text: str,
    sentence_scores: list[dict],
) -> dict:
    """
    Two-signal hybrid:
    - BERT (70%): neural, trained on Wikipedia pairs
    - Stylometrics (30%): model-agnostic structural features
    """
    stylo_result = compute_stylometric_score(text)
    stylo_score = stylo_result["stylometric_ai_score"]

    bert_prob = bert_result["ai_probability"]

    bert_weight  = 0.70
    stylo_weight = 0.30

    final_probability = (
        bert_prob   * bert_weight +
        stylo_score * stylo_weight
    )

    final_probability = round(max(0.0, min(1.0, final_probability)), 4)

    ai_phrases = find_ai_phrases(text)

    explanation = generate_explanation(
        ai_probability=final_probability,
        stylometric_data=stylo_result,
        sentence_scores=sentence_scores,
        ai_phrases=ai_phrases,
    )

    is_ai = final_probability >= settings.CONFIDENCE_THRESHOLD

    return {
        "ai_probability": final_probability,
        "is_ai": is_ai,
        "verdict": explanation["verdict"],
        "confidence_threshold": settings.CONFIDENCE_THRESHOLD,
        "scores": {
            "bert_score": bert_prob,
            "stylometric_score": round(stylo_score, 4),
            "bert_weight": bert_weight,
            "stylometric_weight": stylo_weight,
        },
        "stylometric_features": stylo_result["features"],
        "sentence_scores": sentence_scores,
        "explanation": explanation,
        "chunks_analyzed": bert_result.get("chunks_analyzed", 1),
    }