spec_version: 1
name: counterfeint
description: >
  Three-agent adversarial ad-fraud environment. A Fraudster proposes scam
  ads into a review queue, an Investigator audits each ad with a six-tool
  surface (advertiser_history, landing_page, payment_method, targeting_overlap,
  campaign_structure, policy_classifier — a Llama Guard 3 / Purple Llama
  mock), and an Auditor grades both sides with deterministic Track A
  (reasoning quality) + Track B (plausibility) rubrics. Designed for the
  Multi-Agent Interactions and Self-Improvement hackathon themes.
type: space
sdk: docker
runtime: fastapi
app: server.app:app
port: 8000

# --- Multi-agent / OpenEnv metadata (additive, optional) ------------------
roles:
  - name: fraudster
    description: Proposes/modifies/skips ad creatives to slip fraud past the Investigator.
  - name: investigator
    description: Reviews queued ads, calls six investigation tools, issues verdicts.
  - name: auditor
    description: Oversight layer. Grades Investigator reasoning (Track A) + Fraudster plausibility (Track B). Not trained in Phase 3; retained for evaluation.

themes:
  - multi-agent-interactions
  - self-improving-agents
  - long-horizon-planning

tags:
  - multi-agent
  - adversarial
  - ads
  - trust-and-safety
  - self-play
  - rubric-based
  - llama-guard

# Endpoints exposed for external judges / evaluators (see server/public_api.py).
public_endpoints:
  - method: GET
    path: /api/v1/match/{match_id}
  - method: POST
    path: /api/v1/tools/policy_classifier
  - method: GET
    path: /api/v1/metrics