{
  "schema": "anvil-serving.reasoning-limit-probe-plan/v1",
  "status": "prepared_not_executed",
  "execution_authorized": false,
  "live_identity_verified": false,
  "ignis": "hold",
  "runtime": {
    "engine_revision": "8244d1c44348ae8a5bbfc9f1b607021652283ed4",
    "build_revision": "6a833b698f0b41a9c12922343553eee419b9f465",
    "image_digest": "sha256:745feb916ee816f5409fd58386114be4f22fbede7f9c0e0a7b8c4fe8dac0fb29",
    "model_revision": "76348ddb2be2e35e39ecfc00e75dc07703f4f3d5",
    "template_sha256": "b5bbea09a508f00213386271fae738fb6373bf396e0cc19baaa9675a10952bb0",
    "reasoning_parser": "glm45",
    "model_runner": "V2",
    "speculative_tokens": 3
  },
  "limits": {
    "concurrency": 1,
    "maximum_requests": 2,
    "request_timeout_s": 120,
    "total_window_s": 300,
    "retries": 0
  },
  "isolated_tier_extra_body": {
    "thinking_token_budget": 64,
    "reasoning_effort": "max",
    "temperature": 0.6,
    "top_p": 0.95
  },
  "request_common": {
    "model": "llm.primary",
    "max_tokens": 512,
    "reasoning_effort": "max",
    "temperature": 0.6,
    "top_p": 0.95,
    "stream": false
  },
  "cases": [
    {
      "id": "boundary",
      "prompt": "Solve SEND + MORE = MONEY in decimal. Different letters are different digits and leading digits cannot be zero. Return only the three integers SEND, MORE, MONEY separated by commas.",
      "expected_integers": [
        9567,
        1085,
        10652
      ]
    },
    {
      "id": "natural_end_fallback",
      "condition": "Only if boundary case naturally ends reasoning below 64 tokens with coherent output and enforcement remains untested.",
      "prompt": "Solve CROSS + ROADS = DANGER in decimal. Different letters are different digits and leading digits cannot be zero. Return only the three integers CROSS, ROADS, DANGER separated by commas.",
      "expected_integers": [
        96233,
        62513,
        158746
      ]
    }
  ],
  "required_evidence": [
    "live package and template identity parity",
    "exact forwarded upstream request body",
    "request start and exactly one response or typed transport error",
    "finish reason and nullable native token usage",
    "token IDs or equivalent verified natural versus forced end-marker evidence",
    "visible-answer oracle result separately from boundary result",
    "pre/post RAM, pressure, GPU and cgroup measurements",
    "owned request drain and unchanged incumbent configuration"
  ],
  "stop_conditions": [
    "identity mismatch",
    "requested controls absent or overridden",
    "transport refusal or timeout",
    "length-only reasoning",
    "repeated end markers",
    "incoherent or repetitive continuation",
    "missing boundary evidence",
    "host resource pressure"
  ],
  "execution_prerequisites": [
    "Notify user before any running-model inspection or request.",
    "Create and review an isolated managed request path to the same incumbent endpoint; do not alter shared tier controls.",
    "Verify deployed source/template bytes and assess forced-logit split-top-p risk before inference.",
    "Bind managed raw-response and marker capture; the current Aider request observations alone cannot prove forced closure."
  ]
}