misalignment_monitor: resources_servers: misalignment_monitor: domain: safety description: Layered misalignment monitor for training and evaluation rollouts (activation screen, model review, investigation) value: Catch reward hacking, deception, sandbox escapes and other misaligned agent behavior without changing rewards # The judge is a separate model server: a monitored policy must not grade itself. judge_model_server: type: responses_api_models name: judge_model verified: false misalignment_monitor_simple_agent: responses_api_agents: simple_agent: entrypoint: app.py resources_server: type: resources_servers name: misalignment_monitor model_server: type: responses_api_models name: policy_model datasets: - name: example type: example jsonl_fpath: resources_servers/misalignment_monitor/data/example.jsonl num_repeats: 1 license: Apache 2.0 misalignment_monitor_environment_server: environment_servers: legacy_agent: entrypoint: app.py agent_server: type: responses_api_agents name: misalignment_monitor_simple_agent