Results / results/candidate_01/20260812T052415Z/results.json
{
"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": "2026-08-12T05:24:15.161152+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,
"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": 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": 294.29358902494874,
"fidelity_vs_q8_mean": 0.486328125
},
"pure_q8": {
"tokens_per_s": 204.40615588770143,
"fidelity_vs_q8_mean": 1
}
},
"runs": [
{
"mode": "margin",
"tau": 1,
"window": 32,
"fidelity_vs_q8_mean": 0.5481770833333334,
"tokens_per_s": 110.29541090003036,
"deferral_rate": 0.369140625,
"rollback_rate": 0.06412760416666667,
"sweeps_per_100tok": 8.203125,
"sweep_latency_s_mean": 0.012119657567447055,
"logical_verify_bytes_per_token": 149978260.27734375,
"raw": {
"tokens": 3072,
"deferred": 1134,
"sweeps": 252,
"rollbacks": 197,
"sweep_s": 3.054153706996658,
"gen_s": 24.798319709004318,
"logical_bytes": 460733215572
}
},
{
"mode": "margin",
"tau": 2,
"window": 32,
"fidelity_vs_q8_mean": 0.6611328125,
"tokens_per_s": 78.09292765190605,
"deferral_rate": 0.99609375,
"rollback_rate": 0.10286458333333333,
"sweeps_per_100tok": 11.5234375,
"sweep_latency_s_mean": 0.011960842370043571,
"logical_verify_bytes_per_token": 210683746.58007812,
"raw": {
"tokens": 3072,
"deferred": 3060,
"sweeps": 354,
"rollbacks": 316,
"sweep_s": 4.234138198995424,
"gen_s": 35.10361097200803,
"logical_bytes": 647220469494
}
},
{
"mode": "verify-all",
"tau": 2,
"window": 32,
"fidelity_vs_q8_mean": 0.8863932291666666,
"tokens_per_s": 68.72679721691286,
"deferral_rate": 1.0904947916666667,
"rollback_rate": 0.12076822916666667,
"sweeps_per_100tok": 12.98828125,
"sweep_latency_s_mean": 0.0121271853809379,
"logical_verify_bytes_per_token": 237465578.77246094,
"raw": {
"tokens": 3072,
"deferred": 3350,
"sweeps": 399,
"rollbacks": 371,
"sweep_s": 4.838746966994222,
"gen_s": 39.85997499300447,
"logical_bytes": 729494257989
}
}
],
"quality_bf16_judge": {
"pure_q4": {
"mean_logprob_bf16": -0.3872091167916854,
"n": 24
},
"pure_q8": {
"mean_logprob_bf16": -0.13283097247282663,
"n": 24
},
"margin_tau1.0": {
"mean_logprob_bf16": -0.2670727012058099,
"n": 24
},
"margin_tau2.0": {
"mean_logprob_bf16": -0.17614666558802128,
"n": 24
},
"verify-all_tau2.0": {
"mean_logprob_bf16": -0.12973888901372752,
"n": 24
}
}
}