genre_policy.py

import json
import os

class GenreDecisionPolicy:
    def __init__(self, policy_path="decision_policy.json"):
        with open(policy_path, "r") as f:
            self.policy = json.load(f)
        self.distance_matrix = self.policy.get("distance_matrix", {})
        self.exact_rules = self.policy.get("exact_auto_label", {})
        self.ambig_rules = self.policy.get("ambiguous_bucket", {})

    def get_distance(self, g1, g2):
        if g1 == g2: return 0.0
        if g1 in self.distance_matrix and g2 in self.distance_matrix[g1]:
            return self.distance_matrix[g1][g2]
        if g2 in self.distance_matrix and g1 in self.distance_matrix[g2]:
            return self.distance_matrix[g2][g1]
        return 0.75

    def decide(self, top_predictions):
        if not top_predictions or len(top_predictions) < 2:
            return {
                "decision": "abstain",
                "final_label": None,
                "ambiguity_bucket": None,
                "reason": "not_enough_predictions"
            }
        
        top1 = top_predictions[0][0]
        top1_conf = top_predictions[0][1]
        top2 = top_predictions[1][0]
        top2_conf = top_predictions[1][1]
        
        margin = top1_conf - top2_conf
        dist = self.get_distance(top1, top2)

        decision = "abstain"
        final_label = None
        ambiguity_bucket = None
        reason = "low_confidence"

        # A. Exact Auto Label
        if top1_conf >= self.exact_rules.get("min_top1_confidence", 0.80) and margin >= self.exact_rules.get("min_margin", 0.15):
            decision = "exact_auto_label"
            final_label = top1
            reason = "high_confidence_clear_margin"
        
        # B. Ambiguous Bucket
        elif (top1_conf >= self.ambig_rules.get("min_top1_confidence", 0.50) and 
              top1_conf < self.ambig_rules.get("max_top1_confidence", 0.80) and 
              margin < self.ambig_rules.get("max_margin", 0.20) and 
              dist <= self.ambig_rules.get("max_top1_top2_distance", 0.25)):
            decision = "ambiguous_bucket"
            pair = sorted([top1, top2])
            s1 = pair[0].replace(" / ", "/").replace("Fiction", "Fic")
            s2 = pair[1].replace(" / ", "/").replace("Fiction", "Fic")
            ambiguity_bucket = f"{s1} / {s2} Ambiguous"
            final_label = ambiguity_bucket
            reason = "semantic_proximity_low_margin"

        # C. Abstain Reasons
        elif top1_conf < self.ambig_rules.get("min_top1_confidence", 0.50):
            reason = "low_signal"
        elif margin < self.exact_rules.get("min_margin", 0.15) and dist >= 0.60:
            reason = "semantic_conflict"
        else:
            reason = "insufficient_certainty"

        return {
            "decision": decision,
            "final_label": final_label,
            "ambiguity_bucket": ambiguity_bucket,
            "reason": reason
        }