{
  "backbone": {
    "repo_id": "Qwen/Qwen3-32B",
    "revision": "9216db5781bf21249d130ec9da846c4624c16137"
  },
  "code_adapter": {
    "rank": 32,
    "alpha": 32,
    "target_modules": [
      "q_proj",
      "k_proj",
      "v_proj",
      "o_proj",
      "gate_proj",
      "up_proj",
      "down_proj"
    ],
    "identity_files": "audit_support/<variant>/<arm>/adapters.json"
  },
  "control": "Exact zero-delta code adapter with the same tensor shapes and configuration; same lora-code route.",
  "server": {
    "tensor_parallel_size": 1,
    "context_length": 32768,
    "maximum_lora_rank": 64,
    "max_running_requests": 2,
    "production_task_concurrency_per_arm": 1,
    "seed": 42,
    "deterministic_inference": true,
    "radix_cache": false,
    "cuda_graph": false
  },
  "sampling": {
    "temperature": 0.0,
    "top_k": 20,
    "repetition_penalty": 1.05
  },
  "scope": "Deployment identifiers and adapter hashes, not a distribution of model weights. Environment files describe the handoff software stack; consult requests/protocols for each historical variant."
}