core/aggregator.py

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

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


def aggregate_scores(
    bert_result: dict,
    text: str,
    sentence_scores: list[dict],
) -> dict:
    """
    Three-signal hybrid:
    - BERT (60%): neural classifier fine-tuned on AI vs human text
    - Perplexity (20%): GPT-2 perplexity — model-agnostic, works for any LLM
    - Stylometrics (20%): structural writing features

    Perplexity replaces pure BERT weight because it catches AI text that
    slop-detector-bert wasn't trained on (modern GPT-4/Claude output).
    """
    stylo_result = compute_stylometric_score(text)
    stylo_score = stylo_result["stylometric_ai_score"]

    bert_prob = bert_result["ai_probability"]

    # Perplexity signal — lower perplexity = more predictable = more AI-like
    raw_perplexity = perplexity_scorer.get_perplexity(text)
    perplexity_score = perplexity_scorer.perplexity_to_ai_score(raw_perplexity)

    bert_weight       = 0.60
    perplexity_weight = 0.20
    stylo_weight      = 0.20

    final_probability = (
        bert_prob        * bert_weight +
        perplexity_score * perplexity_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,
        perplexity=raw_perplexity,
    )

    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),
        "perplexity": round(raw_perplexity, 2) if raw_perplexity is not None else None,
        "perplexity_score": round(perplexity_score, 4),
    }