Set or clear the org-level llm_judge model default.
model_settings=null clears the override (back to the code role default). A non-null entry must reference a model in MODEL_REGISTRY (422 otherwise); since the catalog == registry, anything the UI dropdown offers passes. The ModelChainEntry shape also validates the config knobs at the boundary.
Successful Response
The org-level default model for llm_judge policy checks.
model_settings is the org's chosen ModelChainEntry ({model, config}), or null when unset — in which case the runtime falls back to the code role default (gpt-5.4-mini / effort low). The UI renders that fallback label from the model catalog's policy_judge role default, so this response carries ONLY the override.