{
  "experiment": "candidate_01_deferred_refinement",
  "author": "Simon-Pierre Boucher",
  "contact": "contact@spboucher.ai",
  "manifest": {
    "author": "Simon-Pierre Boucher",
    "contact": "contact@spboucher.ai",
    "project": "localvm-research",
    "collected_utc": "2026-08-12T05:15:01.339960+00:00",
    "chip": {
      "brand": "Apple M5 Max",
      "arch": "arm64",
      "cores_total": 18,
      "cores_performance": 6,
      "cores_efficiency": 12,
      "gpu_cores": 40
    },
    "memory": {
      "unified_bytes": 51539607552,
      "unified_gb": 48.0,
      "pagesize": 16384
    },
    "ssd": {
      "model": "APPLE SSD AP2048Z",
      "size": "2 TB",
      "smart_status": "Verified"
    },
    "os": {
      "product": "macOS",
      "version": "27.0",
      "build": "26A5388g",
      "kernel": "27.0.0"
    },
    "software": {
      "python": "3.14.4",
      "mlx": "0.32.0",
      "mlx_lm": "0.31.3",
      "torch": null,
      "numpy": "2.5.2"
    },
    "git": {
      "commit": "bfeefb4674459d923b94e794f3f6f971cecf4570",
      "dirty_tree": true
    },
    "thermal_level_at_collect": null
  },
  "config": {
    "build": false,
    "per_domain": 4,
    "max_tokens": 128,
    "window": 32,
    "taus": "1.0,2.0"
  },
  "model": "mlx-community/Qwen3-1.7B-bf16",
  "q4_resident_bytes": 968080210,
  "q8_stream_bytes": 1828306411,
  "baselines": {
    "pure_q4": {
      "tokens_per_s": 287.5061867410425,
      "fidelity_vs_q8_mean": 0.486328125
    },
    "pure_q8": {
      "tokens_per_s": 200.05224940165664,
      "fidelity_vs_q8_mean": 1.0
    }
  },
  "runs": [
    {
      "mode": "margin",
      "tau": 1.0,
      "window": 32,
      "fidelity_vs_q8_mean": 0.5517578125,
      "tokens_per_s": 99.5551573243222,
      "deferral_rate": 0.3802083333333333,
      "rollback_rate": 0.06673177083333333,
      "sweeps_per_100tok": 8.3984375,
      "sweep_latency_s_mean": 0.021746737585273467,
      "logical_verify_bytes_per_token": 153549171.23632812,
      "raw": {
        "tokens": 3072,
        "deferred": 1168,
        "sweeps": 258,
        "rollbacks": 205,
        "sweep_s": 5.610658297000555,
        "gen_s": 25.246607992000463,
        "logical_bytes": 471703054038
      }
    },
    {
      "mode": "margin",
      "tau": 2.0,
      "window": 32,
      "fidelity_vs_q8_mean": 0.6484375,
      "tokens_per_s": 71.1569560818351,
      "deferral_rate": 1.0032552083333333,
      "rollback_rate": 0.1064453125,
      "sweeps_per_100tok": 11.81640625,
      "sweep_latency_s_mean": 0.022036683184571613,
      "logical_verify_bytes_per_token": 216040113.0185547,
      "raw": {
        "tokens": 3072,
        "deferred": 3082,
        "sweeps": 363,
        "rollbacks": 327,
        "sweep_s": 7.9993159959994955,
        "gen_s": 35.17285112799891,
        "logical_bytes": 663675227193
      }
    },
    {
      "mode": "verify-all",
      "tau": 2.0,
      "window": 32,
      "fidelity_vs_q8_mean": 0.8824869791666666,
      "tokens_per_s": 63.58000089353261,
      "deferral_rate": 1.0387369791666667,
      "rollback_rate": 0.11653645833333333,
      "sweeps_per_100tok": 12.630208333333334,
      "sweep_latency_s_mean": 0.023310723796370926,
      "logical_verify_bytes_per_token": 230918908.6809896,
      "raw": {
        "tokens": 3072,
        "deferred": 3191,
        "sweeps": 388,
        "rollbacks": 358,
        "sweep_s": 9.04456083299192,
        "gen_s": 39.27251933100797,
        "logical_bytes": 709382887468
      }
    }
  ]
}