openenv.yaml

name: shadowops
version: 1.0.0
description: >
  ShadowOps is a cybersecurity incident-response RL environment where agents
  decide whether cloud, CI, IAM, network, and pentest actions should be allowed,
  blocked, forked to human review, or quarantined.
entrypoint: backend-ml/openenv_shadowops_env.py:ShadowOpsOpenEnvV1
schema_contract: backend-ml/schema_contract.json
interface:
  reset: reset()
  step: step(action)
  state: state()
  close: close()
actions:
  - ALLOW
  - BLOCK
  - QUARANTINE
  - FORK
observation:
  type: object
  fields:
    prompt: Cybersecurity incident prompt for the agent.
    risk_vector: Sixteen-dimensional risk feature vector.
    incident_state: Production, memory, health, and quarantine state.
    available_actions: Valid supervisor actions.
reward:
  type: composable
  rubrics:
    - correct_action_reward
    - safety_reward
    - false_positive_penalty
    - missing_evidence_penalty
    - risk_calibration_reward
    - memory_chain_reward
    - safe_outcome_reward
    - invalid_output_penalty
done:
  condition: episode step count reaches episode_max_length.
safe_defaults:
  model_free: true
  deterministic_seed: 42
  no_huggingface_jobs: true