[
  {
    "backend": "numpy",
    "dataset": "generated domain-conditioned linear classification; disjoint RNG draws",
    "evidence_type": "synthetic_cpu_algorithmic_fixture",
    "limitations": [
      "domain label is given to request router",
      "tiny linear experts not transformers",
      "weight bytes exclude attention, KV cache and process memory",
      "timings are noisy CPU microbenchmarks",
      "no real checkpoint weights downloaded or model runtime certified"
    ],
    "machine": "x86_64",
    "not_measured": [
      "real LLM quality",
      "GPU inference",
      "energy",
      "wireless bandwidth",
      "production routing"
    ],
    "numpy_version": "2.3.5",
    "parent": {
      "accuracy": 0.91455078125,
      "batch32_latency_ms_p50": 0.010466,
      "batch32_latency_ms_p95": 0.01125,
      "per_domain_accuracy": {
        "0": 0.9186046511627907,
        "1": 0.9187562688064193,
        "2": 0.9148514851485149,
        "3": 0.9063386944181646
      },
      "retained_experts": 8,
      "router_and_expert_weight_bytes_per_specialist": 5120
    },
    "python": "3.12.13",
    "random": {
      "accuracy": 0.79248046875,
      "batch32_latency_ms_p50": 0.009835,
      "batch32_latency_ms_p95": 0.01095,
      "per_domain_accuracy": {
        "0": 0.44573643410852715,
        "1": 0.9317953861584755,
        "2": 0.9,
        "3": 0.8968779564806055
      },
      "retained_experts": 2,
      "router_and_expert_weight_bytes_per_specialist": 1280
    },
    "random_masks": {
      "0": [
        4,
        5
      ],
      "1": [
        1,
        3
      ],
      "2": [
        1,
        5
      ],
      "3": [
        2,
        6
      ]
    },
    "request_router": "oracle domain labels; NOT a learned production router",
    "schema_version": 1,
    "seed": 7,
    "selected": {
      "accuracy": 0.91455078125,
      "batch32_latency_ms_p50": 0.009844499999999999,
      "batch32_latency_ms_p95": 0.010235049999999999,
      "per_domain_accuracy": {
        "0": 0.9186046511627907,
        "1": 0.9187562688064193,
        "2": 0.9148514851485149,
        "3": 0.9063386944181646
      },
      "retained_experts": 2,
      "router_and_expert_weight_bytes_per_specialist": 1280
    },
    "selected_accuracy_delta_pp": 0.0,
    "selected_masks": {
      "0": [
        1,
        0
      ],
      "1": [
        3,
        2
      ],
      "2": [
        4,
        5
      ],
      "3": [
        6,
        7
      ]
    },
    "selected_weight_reduction_percent": 75.0,
    "selection": "top-2 calibration routing frequency, NOT causal expert importance",
    "splits": {
      "calibration": 1024,
      "heldout": 4096,
      "train": 8192
    },
    "token_top_k": 1
  },
  {
    "backend": "numpy",
    "dataset": "generated domain-conditioned linear classification; disjoint RNG draws",
    "evidence_type": "synthetic_cpu_algorithmic_fixture",
    "limitations": [
      "domain label is given to request router",
      "tiny linear experts not transformers",
      "weight bytes exclude attention, KV cache and process memory",
      "timings are noisy CPU microbenchmarks",
      "no real checkpoint weights downloaded or model runtime certified"
    ],
    "machine": "x86_64",
    "not_measured": [
      "real LLM quality",
      "GPU inference",
      "energy",
      "wireless bandwidth",
      "production routing"
    ],
    "numpy_version": "2.3.5",
    "parent": {
      "accuracy": 0.914794921875,
      "batch32_latency_ms_p50": 0.0103605,
      "batch32_latency_ms_p95": 0.011080999999999999,
      "per_domain_accuracy": {
        "0": 0.9031281533804238,
        "1": 0.9239653512993262,
        "2": 0.9254302103250478,
        "3": 0.9058823529411765
      },
      "retained_experts": 8,
      "router_and_expert_weight_bytes_per_specialist": 5120
    },
    "python": "3.12.13",
    "random": {
      "accuracy": 0.59033203125,
      "batch32_latency_ms_p50": 0.009855,
      "batch32_latency_ms_p95": 0.010469499999999996,
      "per_domain_accuracy": {
        "0": 0.2785065590312815,
        "1": 0.8835418671799807,
        "2": 0.9254302103250478,
        "3": 0.25098039215686274
      },
      "retained_experts": 2,
      "router_and_expert_weight_bytes_per_specialist": 1280
    },
    "random_masks": {
      "0": [
        2,
        7
      ],
      "1": [
        1,
        2
      ],
      "2": [
        4,
        5
      ],
      "3": [
        0,
        1
      ]
    },
    "request_router": "oracle domain labels; NOT a learned production router",
    "schema_version": 1,
    "seed": 19,
    "selected": {
      "accuracy": 0.914794921875,
      "batch32_latency_ms_p50": 0.009835,
      "batch32_latency_ms_p95": 0.010269499999999999,
      "per_domain_accuracy": {
        "0": 0.9031281533804238,
        "1": 0.9239653512993262,
        "2": 0.9254302103250478,
        "3": 0.9058823529411765
      },
      "retained_experts": 2,
      "router_and_expert_weight_bytes_per_specialist": 1280
    },
    "selected_accuracy_delta_pp": 0.0,
    "selected_masks": {
      "0": [
        0,
        1
      ],
      "1": [
        3,
        2
      ],
      "2": [
        5,
        4
      ],
      "3": [
        6,
        7
      ]
    },
    "selected_weight_reduction_percent": 75.0,
    "selection": "top-2 calibration routing frequency, NOT causal expert importance",
    "splits": {
      "calibration": 1024,
      "heldout": 4096,
      "train": 8192
    },
    "token_top_k": 1
  },
  {
    "backend": "numpy",
    "dataset": "generated domain-conditioned linear classification; disjoint RNG draws",
    "evidence_type": "synthetic_cpu_algorithmic_fixture",
    "limitations": [
      "domain label is given to request router",
      "tiny linear experts not transformers",
      "weight bytes exclude attention, KV cache and process memory",
      "timings are noisy CPU microbenchmarks",
      "no real checkpoint weights downloaded or model runtime certified"
    ],
    "machine": "x86_64",
    "not_measured": [
      "real LLM quality",
      "GPU inference",
      "energy",
      "wireless bandwidth",
      "production routing"
    ],
    "numpy_version": "2.3.5",
    "parent": {
      "accuracy": 0.927490234375,
      "batch32_latency_ms_p50": 0.010396,
      "batch32_latency_ms_p95": 0.012020499999999993,
      "per_domain_accuracy": {
        "0": 0.9103182256509161,
        "1": 0.939712918660287,
        "2": 0.9409448818897638,
        "3": 0.9188376753507014
      },
      "retained_experts": 8,
      "router_and_expert_weight_bytes_per_specialist": 5120
    },
    "python": "3.12.13",
    "random": {
      "accuracy": 0.56298828125,
      "batch32_latency_ms_p50": 0.009905,
      "batch32_latency_ms_p95": 0.010430499999999999,
      "per_domain_accuracy": {
        "0": 0.24011571841851495,
        "1": 0.21052631578947367,
        "2": 0.9360236220472441,
        "3": 0.8877755511022044
      },
      "retained_experts": 2,
      "router_and_expert_weight_bytes_per_specialist": 1280
    },
    "random_masks": {
      "0": [
        2,
        6
      ],
      "1": [
        1,
        7
      ],
      "2": [
        5,
        6
      ],
      "3": [
        2,
        6
      ]
    },
    "request_router": "oracle domain labels; NOT a learned production router",
    "schema_version": 1,
    "seed": 41,
    "selected": {
      "accuracy": 0.927490234375,
      "batch32_latency_ms_p50": 0.0098645,
      "batch32_latency_ms_p95": 0.010875049999999994,
      "per_domain_accuracy": {
        "0": 0.9103182256509161,
        "1": 0.939712918660287,
        "2": 0.9409448818897638,
        "3": 0.9188376753507014
      },
      "retained_experts": 2,
      "router_and_expert_weight_bytes_per_specialist": 1280
    },
    "selected_accuracy_delta_pp": 0.0,
    "selected_masks": {
      "0": [
        1,
        0
      ],
      "1": [
        2,
        3
      ],
      "2": [
        5,
        4
      ],
      "3": [
        7,
        6
      ]
    },
    "selected_weight_reduction_percent": 75.0,
    "selection": "top-2 calibration routing frequency, NOT causal expert importance",
    "splits": {
      "calibration": 1024,
      "heldout": 4096,
      "train": 8192
    },
    "token_top_k": 1
  },
  {
    "backend": "numpy",
    "dataset": "generated domain-conditioned linear classification; disjoint RNG draws",
    "evidence_type": "synthetic_cpu_algorithmic_fixture",
    "limitations": [
      "domain label is given to request router",
      "tiny linear experts not transformers",
      "weight bytes exclude attention, KV cache and process memory",
      "timings are noisy CPU microbenchmarks",
      "no real checkpoint weights downloaded or model runtime certified"
    ],
    "machine": "x86_64",
    "not_measured": [
      "real LLM quality",
      "GPU inference",
      "energy",
      "wireless bandwidth",
      "production routing"
    ],
    "numpy_version": "2.3.5",
    "parent": {
      "accuracy": 0.93359375,
      "batch32_latency_ms_p50": 0.010385499999999999,
      "batch32_latency_ms_p95": 0.011011549999999998,
      "per_domain_accuracy": {
        "0": 0.9324452901998097,
        "1": 0.9429695181907571,
        "2": 0.9343283582089552,
        "3": 0.9247311827956989
      },
      "retained_experts": 8,
      "router_and_expert_weight_bytes_per_specialist": 5120
    },
    "python": "3.12.13",
    "random": {
      "accuracy": 0.51220703125,
      "batch32_latency_ms_p50": 0.009915,
      "batch32_latency_ms_p95": 0.010291499999999999,
      "per_domain_accuracy": {
        "0": 0.4005708848715509,
        "1": 0.9252704031465093,
        "2": 0.3383084577114428,
        "3": 0.3870967741935484
      },
      "retained_experts": 2,
      "router_and_expert_weight_bytes_per_specialist": 1280
    },
    "random_masks": {
      "0": [
        3,
        7
      ],
      "1": [
        1,
        2
      ],
      "2": [
        1,
        7
      ],
      "3": [
        1,
        2
      ]
    },
    "request_router": "oracle domain labels; NOT a learned production router",
    "schema_version": 1,
    "seed": 67,
    "selected": {
      "accuracy": 0.93359375,
      "batch32_latency_ms_p50": 0.009895,
      "batch32_latency_ms_p95": 0.010265999999999999,
      "per_domain_accuracy": {
        "0": 0.9324452901998097,
        "1": 0.9429695181907571,
        "2": 0.9343283582089552,
        "3": 0.9247311827956989
      },
      "retained_experts": 2,
      "router_and_expert_weight_bytes_per_specialist": 1280
    },
    "selected_accuracy_delta_pp": 0.0,
    "selected_masks": {
      "0": [
        0,
        1
      ],
      "1": [
        2,
        3
      ],
      "2": [
        5,
        4
      ],
      "3": [
        7,
        6
      ]
    },
    "selected_weight_reduction_percent": 75.0,
    "selection": "top-2 calibration routing frequency, NOT causal expert importance",
    "splits": {
      "calibration": 1024,
      "heldout": 4096,
      "train": 8192
    },
    "token_top_k": 1
  },
  {
    "backend": "numpy",
    "dataset": "generated domain-conditioned linear classification; disjoint RNG draws",
    "evidence_type": "synthetic_cpu_algorithmic_fixture",
    "limitations": [
      "domain label is given to request router",
      "tiny linear experts not transformers",
      "weight bytes exclude attention, KV cache and process memory",
      "timings are noisy CPU microbenchmarks",
      "no real checkpoint weights downloaded or model runtime certified"
    ],
    "machine": "x86_64",
    "not_measured": [
      "real LLM quality",
      "GPU inference",
      "energy",
      "wireless bandwidth",
      "production routing"
    ],
    "numpy_version": "2.3.5",
    "parent": {
      "accuracy": 0.919677734375,
      "batch32_latency_ms_p50": 0.010386,
      "batch32_latency_ms_p95": 0.011141999999999997,
      "per_domain_accuracy": {
        "0": 0.9089968976215098,
        "1": 0.9160447761194029,
        "2": 0.9196515004840271,
        "3": 0.93359375
      },
      "retained_experts": 8,
      "router_and_expert_weight_bytes_per_specialist": 5120
    },
    "python": "3.12.13",
    "random": {
      "accuracy": 0.921630859375,
      "batch32_latency_ms_p50": 0.009845,
      "batch32_latency_ms_p95": 0.010095999999999999,
      "per_domain_accuracy": {
        "0": 0.905894519131334,
        "1": 0.9384328358208955,
        "2": 0.9070667957405615,
        "3": 0.93359375
      },
      "retained_experts": 2,
      "router_and_expert_weight_bytes_per_specialist": 1280
    },
    "random_masks": {
      "0": [
        0,
        6
      ],
      "1": [
        3,
        4
      ],
      "2": [
        3,
        4
      ],
      "3": [
        6,
        7
      ]
    },
    "request_router": "oracle domain labels; NOT a learned production router",
    "schema_version": 1,
    "seed": 101,
    "selected": {
      "accuracy": 0.919677734375,
      "batch32_latency_ms_p50": 0.009885,
      "batch32_latency_ms_p95": 0.010188499999999998,
      "per_domain_accuracy": {
        "0": 0.9089968976215098,
        "1": 0.9160447761194029,
        "2": 0.9196515004840271,
        "3": 0.93359375
      },
      "retained_experts": 2,
      "router_and_expert_weight_bytes_per_specialist": 1280
    },
    "selected_accuracy_delta_pp": 0.0,
    "selected_masks": {
      "0": [
        1,
        0
      ],
      "1": [
        3,
        2
      ],
      "2": [
        5,
        4
      ],
      "3": [
        7,
        6
      ]
    },
    "selected_weight_reduction_percent": 75.0,
    "selection": "top-2 calibration routing frequency, NOT causal expert importance",
    "splits": {
      "calibration": 1024,
      "heldout": 4096,
      "train": 8192
    },
    "token_top_k": 1
  }
]