name: graphstrike
version: "1.0.0"
description: >
  RL environment for detecting coordinated fake account rings in social networks.
  An AI detective agent must identify all 10 members of a coordinated fake
  account gang before they evade detection. Three difficulty levels with
  increasing network size, decoys, and real-time evasion events.
author: OpenENV Hackathon
tags:
  - social-network
  - fraud-detection
  - graph
  - rl
  - fake-account-detection

# ── Server configuration ──────────────────────────────────────────────────────
server:
  dockerfile: Dockerfile
  port: 7860

environment_variables:
  PORT: "7860"
  HOST: "0.0.0.0"
  WORKERS: "1"

# ── Task definitions (each has a grader) ─────────────────────────────────────
tasks:
  - name: easy
    description: "50 accounts, 10 fakes, no evasion, 30 steps"
    max_steps: 30
    grader:
      endpoint: /grader
      score_range: [0.0, 1.0]
      win_threshold: 0.815
      win_conditions:
        recall: 0.8
        precision: 0.7
      formula: >
        if recall >= 0.8 and precision >= 0.7:
          score = 0.55 + 0.20*recall + 0.15*precision + 0.10*efficiency
        else:
          score = 0.30*recall + 0.10*precision

  - name: medium
    description: "200 accounts, 10 fakes + 20 decoys, 1 evasion event, 50 steps"
    max_steps: 50
    grader:
      endpoint: /grader
      score_range: [0.0, 1.0]
      win_threshold: 0.815
      win_conditions:
        recall: 0.8
        precision: 0.7
      formula: >
        if recall >= 0.8 and precision >= 0.7:
          score = 0.55 + 0.20*recall + 0.15*precision + 0.10*efficiency
        else:
          score = 0.30*recall + 0.10*precision

  - name: hard
    description: "1000 accounts, 10 fakes + 50 decoys, 4 evasion events, 80 steps"
    max_steps: 80
    grader:
      endpoint: /grader
      score_range: [0.0, 1.0]
      win_threshold: 0.868
      win_conditions:
        recall: 0.9
        precision: 0.8
      formula: >
        if recall >= 0.9 and precision >= 0.8:
          score = 0.55 + 0.20*recall + 0.15*precision + 0.10*efficiency
        else:
          score = 0.30*recall + 0.10*precision

# ── Action schema ─────────────────────────────────────────────────────────────
action_schema:
  action_type:
    type: string
    enum: [inspect, investigate_network, flag, unflag, submit]
  account_id:
    type: string
    description: "Required for all actions except submit"

# ── Score range ───────────────────────────────────────────────────────────────
score_range: [0.0, 1.0]
grader_formula: >
  if recall >= win_recall and precision >= win_precision:
    score = 0.55 + 0.20*recall + 0.15*precision + 0.10*efficiency
  else:
    score = 0.30*recall + 0.10*precision