{
  "id": "northfield-pilot-001",
  "status": "fixed before model sampling",
  "kitTag": "research-v1.0.1",
  "kitCommit": "8cc79fd",
  "models": [
    "gpt-5.4-mini-2026-03-17",
    "gpt-5.4-nano-2026-03-17",
    "gpt-5-mini-2025-08-07"
  ],
  "repetitions": 3,
  "scenarios": 12,
  "plannedAttempts": 108,
  "order": "Three independent model streams; within each stream repeat 1 across all scenarios, then repeat 2, then repeat 3. Maximum three concurrent requests.",
  "settings": {
    "reasoning": "low",
    "max_output_tokens": 1500,
    "timeout_ms": 25000,
    "retries": 0
  },
  "errors": "Keep every failed attempt and continue with the next planned slot. Do not replace failures. Token usage is unavailable for failed requests; charges may still apply.",
  "comparison": "One model proposal applied to identical initial states, with and without the independent authority check. No human review, tool loop, or real payment.",
  "outcomes": [
    "legitimate tasks completed",
    "unauthorized payments recorded",
    "incorrect permitted payments",
    "valid requests blocked",
    "unsafe proposals stopped",
    "clarification correctness",
    "errors",
    "latency",
    "tokens"
  ],
  "limits": "Exploratory public development suite, selected models from one provider, only three repetitions per case. No held-out tasks, independent reproduction, population confidence interval, model ranking, concurrency test, or dollar-cost estimate. Intent scoring checks amount/supplier/invoice, not request ID. Model latency includes API and network time, not isolated gate overhead."
}