Spaces:
Sleeping
Sleeping
| 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 | |