{
  "best_depth": 1,
  "model_types": [
    "qwen3_5"
  ],
  "artifact": "Qwen3.8-27B-JANG_2D",
  "quantization_mode": "affine",
  "quantization_bits": 2,
  "note": "best_depth=1, MEASURED against a WARM KV CACHE. 48.49 tok/s vs 48.49 at depth 1 (1.00x). Depth 3 = 39.11 tok/s, a net LOSS. Each extra verified token adds ~20% to the target decode step, so deeper speculation stops paying quickly. output_equivalent NOT asserted.",
  "reason": "stamped by stamp_qwen38_27b; depth sweep + warm-cache wall-clock, 2026-08-24",
  "measured_depth1_acceptance_pct": 85.89,
  "measured_best_depth": 1,
  "measured_depth_curve": [
    {
      "depth": 1,
      "per_position_pct": 85.89,
      "chained_pct": 85.89,
      "expected_tokens_per_cycle": 1.859
    },
    {
      "depth": 2,
      "per_position_pct": 46.37,
      "chained_pct": 45.56,
      "expected_tokens_per_cycle": 2.315
    },
    {
      "depth": 3,
      "per_position_pct": 21.77,
      "chained_pct": 16.13,
      "expected_tokens_per_cycle": 2.476
    }
  ],
  "validated": true,
  "baseline_tok_s": 48.49,
  "best_tok_s": 48.49,
  "speedup_vs_baseline": 1.0,
  "measured_tok_s_by_depth": {
    "1": 48.49,
    "2": 46.13,
    "3": 39.11
  },
  "blocked": false
}