LamChingFung-2425-Anti-FraudX

Feature: Auto Mode (Simulation)

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

Overview

Runs a fully automated AI-vs-AI simulation with no human player input. All three agents (Scammer, Expert, Victim) respond autonomously. Used for training data generation, regression testing, and demonstration purposes.

Flow

AutoModeScene.ts
  │
  │  WebSocket connect to /ws/simulate
  ▼
SimulationRunner.run_simulation(config)
  │
  │  For each round (max 15):
  │    asyncio.gather(
  │      ScammerAgent.generate(victim_last_message),
  │      ExpertAgent.generate(victim_last_message),
  │      VictimAgent.generate(scammer_response + expert_response)
  │    )
  │    → Update trust values
  │    → Check outcome
  │    → Broadcast round event via WebSocket
  │
  │  On outcome or max rounds:
  │    → RecorderAgent.analyse(session)
  │    → Broadcast final result
  ▼
AutoModeScene receives events, renders live simulation

WebSocket Events

All events sent from backend to frontend over /ws/simulate:

Event type Payload Description
round_start { round: number } New round beginning
agent_response { agent, text, trust_changes } Single agent response
trust_update { scammer, expert, alertness } Updated trust values
outcome { winner, reason } Simulation ended
report { analysis: string } RecorderAgent report

Configuration

# SimulationRunner accepts:
config = {
    "scam_type": "investment",
    "victim_persona": "elderly",
    "max_rounds": 15,
    "mode": "auto"
}

Use Cases

Use Case How
Training data generation Run many simulations; conversations saved as JSONL for fine-tuning
Regression testing Verify trust outcome logic after code changes
Demo / presentation Show live AI conversation without requiring a human player
Prompt evaluation Compare prompt versions by running simulations with each

Relevant Files

rpg-platform-v2/src/scenes/AutoModeScene.ts   Frontend WebSocket client + live display
backend/api/simulation_routes.py               WebSocket endpoint /ws/simulate
backend/services/simulation_runner.py          Core simulation loop
backend/agents/victim.py                       VictimAgent (auto-responds)