LamChingFung-2425-Anti-FraudX

Feature: Adaptive Scoring System

Category: Secondary Version: 4.1 Last Updated: 2026-03-11

Overview

Instead of applying fixed trust deltas per round, the Adaptive Scoring System adjusts evaluation weights based on the victim’s persona. An elderly victim is more affected by empathetic appeals; an overconfident victim responds more to hard evidence. This makes the simulation more realistic and psychologically grounded.

Implementation

File: backend/utils/adaptive_scoring.py

WeightConfig

@dataclass
class WeightConfig:
    empathy: float        # How much empathetic language affects trust
    clarity: float        # How much clear, simple explanation affects trust
    evidence: float       # How much factual evidence/citations affect trust
    actionability: float  # How much actionable advice affects trust
    timing: float = 0.0   # Bonus for well-timed intervention

AdaptiveWeightOptimizer

class AdaptiveWeightOptimizer:
    def get_weights(self, persona: str) -> WeightConfig:
        # Returns persona-specific WeightConfig
        ...

    def apply_weights(self, raw_score: float, response_features: dict, persona: str) -> float:
        # Applies weights to raw trust delta
        ...

Persona Weight Profiles

Persona empathy clarity evidence actionability Notes
elderly 0.40 0.35 0.15 0.10 Emotional appeals most effective
average 0.25 0.25 0.25 0.25 Balanced
overconfident 0.10 0.20 0.45 0.25 Needs hard facts to change mind
student 0.20 0.35 0.25 0.20 Responds to clear explanations

Integration with Trust System

Agent response received
  │
  ▼
Response features extracted
  (empathy_score, clarity_score, evidence_score, actionability_score)
  │
  ▼
AdaptiveWeightOptimizer.apply_weights(raw_delta, features, persona)
  │
  ▼
Adjusted trust delta applied to VictimTrustState

This runs inside the trust update pipeline in performance_tracker.py.

Relevant Files

backend/utils/adaptive_scoring.py       AdaptiveWeightOptimizer + WeightConfig
backend/utils/performance_tracker.py    Calls adaptive scoring in trust update
backend/config.py                       PersonaConfig (persona definitions)