{
  "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": "2026-08-12T05:21:38.625190+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": "ad9db7a3b447837ebb15d9776a5d220a13f00f84",
      "dirty_tree": false
    },
    "thermal_level_at_collect": null
  },
  "config": {
    "build": false,
    "per_domain": 1,
    "max_tokens": 48,
    "window": 32,
    "taus": "2.0"
  },
  "model": "mlx-community/Qwen3-1.7B-bf16",
  "q4_resident_bytes": 968080210,
  "q8_stream_bytes": 1828306411,
  "baselines": {
    "pure_q4": {
      "tokens_per_s": 290.07143667847674,
      "fidelity_vs_q8_mean": 0.7604166666666666
    },
    "pure_q8": {
      "tokens_per_s": 200.84102191894104,
      "fidelity_vs_q8_mean": 1.0
    }
  },
  "runs": [
    {
      "mode": "margin",
      "tau": 2.0,
      "window": 32,
      "fidelity_vs_q8_mean": 0.8472222222222222,
      "tokens_per_s": 140.97349004693962,
      "deferral_rate": 0.3298611111111111,
      "rollback_rate": 0.052083333333333336,
      "sweeps_per_100tok": 8.333333333333334,
      "sweep_latency_s_mean": 0.014723937416685354,
      "logical_verify_bytes_per_token": 152358867.58333334,
      "raw": {
        "tokens": 288,
        "deferred": 95,
        "sweeps": 24,
        "rollbacks": 15,
        "sweep_s": 0.3533744980004485,
        "gen_s": 1.6895627939975384,
        "logical_bytes": 43879353864
      }
    },
    {
      "mode": "verify-all",
      "tau": 2.0,
      "window": 32,
      "fidelity_vs_q8_mean": 0.9722222222222222,
      "tokens_per_s": 124.95923503254495,
      "deferral_rate": 0.4340277777777778,
      "rollback_rate": 0.07291666666666667,
      "sweeps_per_100tok": 9.722222222222221,
      "sweep_latency_s_mean": 0.01140881553575517,
      "logical_verify_bytes_per_token": 177752012.18055555,
      "raw": {
        "tokens": 288,
        "deferred": 125,
        "sweeps": 28,
        "rollbacks": 21,
        "sweep_s": 0.3194468350011448,
        "gen_s": 1.985304789998736,
        "logical_bytes": 51192579508
      }
    }
  ],
  "quality_bf16_judge": {
    "pure_q4": {
      "mean_logprob_bf16": -0.25611507147550583,
      "n": 6
    },
    "pure_q8": {
      "mean_logprob_bf16": -0.09550568337241809,
      "n": 6
    },
    "margin_tau2.0": {
      "mean_logprob_bf16": -0.12508264432350794,
      "n": 6
    },
    "verify-all_tau2.0": {
      "mean_logprob_bf16": -0.10116633027791977,
      "n": 6
    }
  }
}