name: cace_env
version: 0.2.0
description: >
  Cultural Context Arbitration Environment — trains LLMs to make culturally-aware
  content moderation decisions using Meta's Oversight Board rulings as a verifiable
  reward oracle. Three-track reward model. Real social network spread simulation.
  No LLM judge. Fully deterministic reward.

client:
  class_name: CACEEnvClient
  module: cace_env.client

action:
  class_name: CACEAction
  module: cace_env.models

observation:
  class_name: CACEObservation
  module: cace_env.models

default_image: cace-env:latest
spec_version: 1

tags:
  - reinforcement-learning
  - content-moderation
  - multi-agent
  - cultural-context
  - oversight-board
  - grpo
  - rlvr