{
  "artifact": "d8_expert_rank",
  "sprint": "D8 / 41",
  "claim": "Whether OLMoE's 64-expert bank spans 64 functional directions: effective rank and top-25%-expert variance share of every expert's output on the same inputs, per layer.",
  "method": "Forward pre-hook captures the residual stream entering each MoE block on 2,048 real tokens; every expert is run on those inputs; outputs are random-projected (JL) to 8,192 dims, centred, SVD'd; routing top-k captured from the gate for the coactivation measures.",
  "gate": {
    "lives": ">= 0.60 layer-median top-25% variance share",
    "dead": "< 0.40"
  },
  "layer_median_variance_share_top25pct": 0.4092,
  "layer_median_normalized_effective_rank": 0.8915,
  "cost_usd_upper_bound": 0.0053,
  "environment": {
    "gpu_name_ground_truth": "NVIDIA L4",
    "gpu_requested": "L4"
  },
  "model": "allenai/OLMoE-1B-7B-0924",
  "n_experts": 64,
  "experts_per_token": 8,
  "n_layers": 16,
  "tokens": 2048,
  "proj_dim": 8192,
  "gpu_name": "NVIDIA L4",
  "per_layer": [
    {
      "layer": 0,
      "effective_rank": 57.814,
      "normalized_effective_rank": 0.9033,
      "variance_share_top25pct_experts": 0.3923,
      "variance_share_top50pct_experts": 0.6365,
      "singular_values_head": [
        44.541,
        31.498,
        29.511,
        27.885,
        27.663,
        27.04,
        26.521,
        24.711
      ],
      "output_cosine_offdiag_mean": 0.0375,
      "output_cosine_offdiag_max": 0.2158,
      "routing_share_within_hottest_25pct": 0.002,
      "routing_top16_load_share": 0.3799
    },
    {
      "layer": 1,
      "effective_rank": 31.473,
      "normalized_effective_rank": 0.4918,
      "variance_share_top25pct_experts": 0.5821,
      "variance_share_top50pct_experts": 0.7667,
      "singular_values_head": [
        120.669,
        41.382,
        38.818,
        33.639,
        33.523,
        32.987,
        32.156,
        30.147
      ],
      "output_cosine_offdiag_mean": 0.0307,
      "output_cosine_offdiag_max": 0.4532,
      "routing_share_within_hottest_25pct": 0.0,
      "routing_top16_load_share": 0.3679
    },
    {
      "layer": 2,
      "effective_rank": 57.837,
      "normalized_effective_rank": 0.9037,
      "variance_share_top25pct_experts": 0.3979,
      "variance_share_top50pct_experts": 0.6562,
      "singular_values_head": [
        56.711,
        45.016,
        40.92,
        40.186,
        38.975,
        38.098,
        37.399,
        37.363
      ],
      "output_cosine_offdiag_mean": 0.0239,
      "output_cosine_offdiag_max": 0.0902,
      "routing_share_within_hottest_25pct": 0.0005,
      "routing_top16_load_share": 0.4136
    },
    {
      "layer": 3,
      "effective_rank": 58.378,
      "normalized_effective_rank": 0.9122,
      "variance_share_top25pct_experts": 0.3902,
      "variance_share_top50pct_experts": 0.6669,
      "singular_values_head": [
        54.588,
        52.102,
        48.668,
        47.047,
        45.869,
        44.985,
        44.171,
        43.3
      ],
      "output_cosine_offdiag_mean": 0.0168,
      "output_cosine_offdiag_max": 0.0996,
      "routing_share_within_hottest_25pct": 0.0,
      "routing_top16_load_share": 0.4463
    },
    {
      "layer": 4,
      "effective_rank": 55.746,
      "normalized_effective_rank": 0.871,
      "variance_share_top25pct_experts": 0.4107,
      "variance_share_top50pct_experts": 0.6665,
      "singular_values_head": [
        107.208,
        88.896,
        69.313,
        66.142,
        65.222,
        61.171,
        60.053,
        59.578
      ],
      "output_cosine_offdiag_mean": 0.0227,
      "output_cosine_offdiag_max": 0.0886,
      "routing_share_within_hottest_25pct": 0.0,
      "routing_top16_load_share": 0.4412
    },
    {
      "layer": 5,
      "effective_rank": 59.329,
      "normalized_effective_rank": 0.927,
      "variance_share_top25pct_experts": 0.372,
      "variance_share_top50pct_experts": 0.6493,
      "singular_values_head": [
        84.496,
        81.242,
        80.758,
        76.103,
        75.635,
        74.337,
        73.816,
        73.044
      ],
      "output_cosine_offdiag_mean": 0.0261,
      "output_cosine_offdiag_max": 0.1113,
      "routing_share_within_hottest_25pct": 0.0015,
      "routing_top16_load_share": 0.4941
    },
    {
      "layer": 6,
      "effective_rank": 57.887,
      "normalized_effective_rank": 0.9045,
      "variance_share_top25pct_experts": 0.3834,
      "variance_share_top50pct_experts": 0.6423,
      "singular_values_head": [
        140.572,
        96.721,
        90.066,
        86.902,
        84.643,
        83.499,
        82.766,
        81.328
      ],
      "output_cosine_offdiag_mean": 0.0194,
      "output_cosine_offdiag_max": 0.2009,
      "routing_share_within_hottest_25pct": 0.0112,
      "routing_top16_load_share": 0.4805
    },
    {
      "layer": 7,
      "effective_rank": 58.842,
      "normalized_effective_rank": 0.9194,
      "variance_share_top25pct_experts": 0.3746,
      "variance_share_top50pct_experts": 0.6469,
      "singular_values_head": [
        128.382,
        113.028,
        109.672,
        98.882,
        96.237,
        93.229,
        92.901,
        90.429
      ],
      "output_cosine_offdiag_mean": 0.0171,
      "output_cosine_offdiag_max": 0.1356,
      "routing_share_within_hottest_25pct": 0.0068,
      "routing_top16_load_share": 0.4544
    },
    {
      "layer": 8,
      "effective_rank": 59.066,
      "normalized_effective_rank": 0.9229,
      "variance_share_top25pct_experts": 0.3749,
      "variance_share_top50pct_experts": 0.6532,
      "singular_values_head": [
        131.159,
        125.409,
        121.654,
        120.968,
        117.778,
        114.75,
        113.886,
        113.391
      ],
      "output_cosine_offdiag_mean": 0.0162,
      "output_cosine_offdiag_max": 0.0801,
      "routing_share_within_hottest_25pct": 0.0015,
      "routing_top16_load_share": 0.4388
    },
    {
      "layer": 9,
      "effective_rank": 57.022,
      "normalized_effective_rank": 0.891,
      "variance_share_top25pct_experts": 0.4112,
      "variance_share_top50pct_experts": 0.6904,
      "singular_values_head": [
        175.201,
        173.82,
        168.519,
        159.315,
        156.931,
        155.436,
        151.728,
        149.0
      ],
      "output_cosine_offdiag_mean": 0.0142,
      "output_cosine_offdiag_max": 0.1365,
      "routing_share_within_hottest_25pct": 0.0,
      "routing_top16_load_share": 0.4049
    },
    {
      "layer": 10,
      "effective_rank": 57.078,
      "normalized_effective_rank": 0.8919,
      "variance_share_top25pct_experts": 0.4076,
      "variance_share_top50pct_experts": 0.6923,
      "singular_values_head": [
        235.284,
        210.561,
        204.173,
        201.804,
        198.891,
        196.065,
        194.717,
        191.291
      ],
      "output_cosine_offdiag_mean": 0.01,
      "output_cosine_offdiag_max": 0.0759,
      "routing_share_within_hottest_25pct": 0.0029,
      "routing_top16_load_share": 0.4227
    },
    {
      "layer": 11,
      "effective_rank": 55.425,
      "normalized_effective_rank": 0.866,
      "variance_share_top25pct_experts": 0.4273,
      "variance_share_top50pct_experts": 0.7212,
      "singular_values_head": [
        320.285,
        290.401,
        275.19,
        263.12,
        259.823,
        252.798,
        248.9,
        243.801
      ],
      "output_cosine_offdiag_mean": 0.0085,
      "output_cosine_offdiag_max": 0.1065,
      "routing_share_within_hottest_25pct": 0.0078,
      "routing_top16_load_share": 0.4392
    },
    {
      "layer": 12,
      "effective_rank": 54.713,
      "normalized_effective_rank": 0.8549,
      "variance_share_top25pct_experts": 0.4569,
      "variance_share_top50pct_experts": 0.7237,
      "singular_values_head": [
        389.31,
        374.597,
        365.49,
        363.905,
        348.678,
        331.735,
        328.003,
        318.4
      ],
      "output_cosine_offdiag_mean": 0.012,
      "output_cosine_offdiag_max": 0.1029,
      "routing_share_within_hottest_25pct": 0.0068,
      "routing_top16_load_share": 0.4621
    },
    {
      "layer": 13,
      "effective_rank": 54.416,
      "normalized_effective_rank": 0.8502,
      "variance_share_top25pct_experts": 0.4726,
      "variance_share_top50pct_experts": 0.7121,
      "singular_values_head": [
        533.29,
        472.404,
        454.101,
        444.689,
        416.113,
        401.726,
        395.827,
        386.882
      ],
      "output_cosine_offdiag_mean": 0.0092,
      "output_cosine_offdiag_max": 0.0802,
      "routing_share_within_hottest_25pct": 0.0171,
      "routing_top16_load_share": 0.4969
    },
    {
      "layer": 14,
      "effective_rank": 52.898,
      "normalized_effective_rank": 0.8265,
      "variance_share_top25pct_experts": 0.4923,
      "variance_share_top50pct_experts": 0.7355,
      "singular_values_head": [
        675.666,
        630.507,
        593.409,
        583.582,
        562.432,
        545.856,
        532.938,
        530.912
      ],
      "output_cosine_offdiag_mean": 0.0692,
      "output_cosine_offdiag_max": 0.2295,
      "routing_share_within_hottest_25pct": 0.0044,
      "routing_top16_load_share": 0.482
    },
    {
      "layer": 15,
      "effective_rank": 38.576,
      "normalized_effective_rank": 0.6027,
      "variance_share_top25pct_experts": 0.6106,
      "variance_share_top50pct_experts": 0.8073,
      "singular_values_head": [
        1983.943,
        1164.06,
        1093.061,
        954.95,
        838.492,
        817.584,
        790.935,
        777.573
      ],
      "output_cosine_offdiag_mean": 0.449,
      "output_cosine_offdiag_max": 0.78,
      "routing_share_within_hottest_25pct": 0.0103,
      "routing_top16_load_share": 0.5505
    }
  ],
  "wall_s": 24.0,
  "verdict": "BETWEEN THE GATES: top-25% share 0.409, normalized effective rank 0.891. Report, no sprint.",
  "does_not_prove": [
    "One MoE (OLMoE-1B-7B, 64 experts, 8 active), 2,048 prose tokens. Expert rank is architecture- and data-specific; 2607.28308 measured six MoEs and found the premise does not generalise.",
    "Rank of OUTPUTS on shared inputs bounds compressibility from above; it says nothing about task quality after any compression (REAP's axis).",
    "Random projection to 8,192 dims perturbs singular values at the ~1% level.",
    "No weight is modified; no pruning heuristic is built or evaluated."
  ],
  "generated_utc": "2026-09-02T15:32:30.246655+00:00",
  "git_rev": "05bb792",
  "git_dirty": true,
  "run_id": "40440606adcd4950",
  "host_kind": "Darwin-arm64",
  "provenance_schema": "gpu-cert-provenance/1"
}