{
  "measured_at": "2026-09-18T12:55:21.035Z",
  "task": "Approve running a Python script against example.com, as requested by the user.",
  "environment": "One Mac; sequential calls to the deployed Jevgate endpoint.",
  "jevgate": {
    "version": "0.2.1",
    "timing": "Node process spawn through exit, including task lookup, HTTP request, local logging, and decision output. The proposed command was evaluated, not executed.",
    "latency_ms": [598, 473, 374, 432, 544, 382, 396, 449, 419, 465, 446, 367],
    "allowed": 12
  },
  "codex_reviewer": {
    "model": "codex-auto-review",
    "reasoning_effort": "low",
    "timing": "Review start through approval decision, captured in earlier runs of the same network-fetch task on the same Mac.",
    "latency_ms": [4039, 3030],
    "allowed": 2
  },
  "comparison": "Ratio of median approval latencies, rounded down to a whole-number multiple. All captured samples in these runs are included.",
  "limitations": [
    "Early comparison of one routine task, with two recorded Codex reviews and twelve repeated Jevgate hook calls; not a randomized paired benchmark.",
    "Approval latency only; does not establish faster task completion, developer time recovered, or customer dollar savings.",
    "Does not establish equal decision quality or safety. Requests Jevgate cannot approve still incur the usual review after the hook."
  ]
}