{
  "title": "Custodian \u2014 Mistral and Llama Nemotron on confidential H100",
  "date": "2026-09-23",
  "type": "Selected public validation results",
  "environment": {
    "provider": "VoltageGPU",
    "cpu_tee": "Intel TDX",
    "gpu": "NVIDIA H100 80 GB",
    "gpu_confidential_mode": true,
    "inference_engine": "vLLM",
    "wcm_version": "0.28.2"
  },
  "measurement": "Milliseconds from the shortened expiry deadline or injected withdrawal/outage trigger to observed worker exit. Withdrawal includes renewal and evidence checks. Individual observations, not a latency SLA or proof of GPU memory erasure.",
  "configuration": "Public model checkpoints; two in-process holder roles and a loopback broker under one administrative account; stock guest image. This run establishes supervised inference lifecycle behavior in this configuration, not independently administered organizations or a qualified immutable serving image.",
  "scenario_method": {
    "expiry": "Shortened a live lease; this was not a full-duration renewal soak.",
    "withdrawal": "Customer holder withheld approval at renewal.",
    "outage": "Injected broker unavailability on renewal."
  },
  "models": [
    {
      "model": "mistralai/Mistral-7B-Instruct-v0.3",
      "revision": "c170c708c41dac9275d15a8fff4eca08d52bab71",
      "passed": true,
      "scenarios": [
        {
          "scenario": "expiry",
          "observed_worker_exit_ms": 678.39,
          "output_check_passed": true,
          "generation_observed": true,
          "worker_stopped": true,
          "output_stopped_changing": true,
          "unauthorized_restart_refused": true
        },
        {
          "scenario": "holder-withdrawal",
          "observed_worker_exit_ms": 22421.19,
          "output_check_passed": true,
          "generation_observed": true,
          "worker_stopped": true,
          "output_stopped_changing": true,
          "unauthorized_restart_refused": true
        },
        {
          "scenario": "broker-outage",
          "observed_worker_exit_ms": 679.84,
          "output_check_passed": true,
          "generation_observed": true,
          "worker_stopped": true,
          "output_stopped_changing": true,
          "unauthorized_restart_refused": true
        }
      ]
    },
    {
      "model": "nvidia/Llama-3.1-Nemotron-Nano-8B-v1",
      "revision": "54641c1611fcff44fa4865626462445e0a153fc7",
      "passed": true,
      "scenarios": [
        {
          "scenario": "expiry",
          "observed_worker_exit_ms": 688.55,
          "output_check_passed": true,
          "generation_observed": true,
          "worker_stopped": true,
          "output_stopped_changing": true,
          "unauthorized_restart_refused": true
        },
        {
          "scenario": "holder-withdrawal",
          "observed_worker_exit_ms": 22724.21,
          "output_check_passed": true,
          "generation_observed": true,
          "worker_stopped": true,
          "output_stopped_changing": true,
          "unauthorized_restart_refused": true
        },
        {
          "scenario": "broker-outage",
          "observed_worker_exit_ms": 675.08,
          "output_check_passed": true,
          "generation_observed": true,
          "worker_stopped": true,
          "output_stopped_changing": true,
          "unauthorized_restart_refused": true
        }
      ]
    }
  ]
}