outputs/tail_stability_reported.json

{
  "status": "REPORTED-NOT-RUN",
  "reason": "Qwen3-8B L2/L3 CoT rollouts not produced (vLLM/driver ABI mismatch).",
  "tail_reported": {
    "L1 Direct": {
      "p50": 0.63,
      "p90": 2.57,
      "p95": 3.27
    },
    "L1+Tool": {
      "p50": 0.52,
      "p90": 2.47,
      "p95": 4.27
    },
    "L2 CoT": {
      "p50": 0.25,
      "p90": 1.7,
      "p95": 2.01
    },
    "L3 CoT": {
      "p50": 0.35,
      "p90": 1.95,
      "p95": 2.92
    }
  },
  "L2_p95": 2.01,
  "L3_p95": 2.92,
  "L2_tighter_tail": true
}