{
 "_doc": "Every field of the lab's result file for this result, in the file's own order, copied unchanged. The file itself is not served: it names a path on the lab's machine and a former internal project name that collides with a third party's. So a field that carries either, or a word this site keeps for its own process, is committed by its sha256 instead of printed, and says why; no field is dropped and none is edited. The whole file is identified by its own sha256 below.",
 "id": "vllm:N22",
 "source": {
  "lab": "vLLM lab",
  "commit": "eff30fb87201cb1690872532e5585697168543bf",
  "file": "results/multimodel_demand.json",
  "sha256": "460ba7a3e386e12920f9ad98213a999a3f07803274720ad3d7aa73337d2f7b22",
  "bytes": 4496
 },
 "fields": {
  "artifact": "multimodel_demand",
  "sprint": "C29 / C22",
  "claim": {
   "withheld": true,
   "why": "uses a word the published record keeps for its own process (sprint)",
   "sha256_of_value": "f0cbdbfe4f827cecef0b22bb9c083f9ca9dc057225a2a52c7562de0c979f9ad6",
   "bytes": 148,
   "hashed_as": "utf-8 text"
  },
  "method": "Reads the WildChat labelled fixture and multimodel.jsonl, computes the model-share distribution, HHI and Shannon entropy of each, and confirms by key inspection that the Mooncake capture carries no model field.",
  "wildchat_path": {
   "withheld": true,
   "why": "carries a machine path or a private name",
   "sha256_of_value": "d9aed1c7e0eed95c9bf404fab3214dabf2771d8269d5f4349ef6dc80183775f3",
   "bytes": 122,
   "hashed_as": "utf-8 text"
  },
  "real_labelled": {
   "n_models": 2,
   "top1_share": 0.637875,
   "top5_share": 1.0,
   "hhi": 0.538019,
   "entropy_bits": 0.944433,
   "models_below_1pct": 0,
   "shares_top8": [
    0.637875,
    0.362125
   ],
   "n_requests": 8000,
   "n_tenants": 1001,
   "n_sessions": 2592,
   "block_tokens": 16,
   "block_reuse_frac": 0.701471,
   "dataset": "allenai/WildChat-1M",
   "license": "ODC-BY (WildChat-1M); see the dataset card",
   "is_real_trace": false,
   "derived_from_real_text": true
  },
  "self_authored": {
   "n_models": 60,
   "top1_share": 0.244,
   "top5_share": 0.540667,
   "hhi": 0.093309,
   "entropy_bits": 4.55795,
   "models_below_1pct": 42,
   "shares_top8": [
    0.244,
    0.135667,
    0.069,
    0.053,
    0.039,
    0.030333,
    0.026333,
    0.025333
   ],
   "n_requests": 3000
  },
  "mooncake_request_keys": [
   "hash_ids",
   "input_length",
   "output_length",
   "timestamp"
  ],
  "mooncake_has_model_field": false,
  "hhi_ratio_real_over_ours": 5.766,
  "entropy_ratio_ours_over_real": 4.8261,
  "a_real_labelled_multimodel_trace_exists": true,
  "verdict": {
   "withheld": true,
   "why": "uses a word the published record keeps for its own process (sprint)",
   "sha256_of_value": "8fccb92ea7465d730a0b1103404a7069dd086f698b342c98bca3eab18eb58e3a",
   "bytes": 1696,
   "hashed_as": "utf-8 text"
  },
  "does_not_prove": [
   "The trace is DERIVED FROM REAL TEXT, not a real trace. Its own header says so. Model, tenant, session and arrival are real; block ids are chained hashes of the real text under tiktoken cl100k_base, so its 0.7015 reuse is a property of that derivation and NOT a measured KV hit rate.",
   {
    "withheld": true,
    "why": "uses a word the published record keeps for its own process (sprint)",
    "sha256_of_value": "14363e6bc63caac5e3278b043d7b2ddad126c6f767997d9cbc270fec31b1705a",
    "bytes": 184,
    "hashed_as": "utf-8 text"
   },
   "`tenant` is a hashed IP, which the fixture itself flags as a proxy for a person: NAT puts many users behind one endpoint. Per-tenant conclusions inherit that.",
   "WildChat is public chat traffic. A provider's multi-variant serving mix is a different population and nothing here samples it."
  ],
  "cost_usd": 0.0,
  "generated_utc": "2026-08-28T18:58:55.663192+00:00",
  "git_rev": "a8b58f3",
  "git_dirty": true,
  "run_id": "504d5911064d4d6f",
  "host_kind": "Darwin-arm64",
  "provenance_schema": "gpu-cert-provenance/1"
 },
 "display": {
  "real_top1_share_pct": "63.8%",
  "real_second_share_pct": "36.2%",
  "self_authored_top1_share_pct": "24.4%",
  "real_n_requests": "8,000",
  "self_authored_n_requests": "3,000"
 },
 "ratio_check": {
  "hhi_real_over_self_authored": 5.766,
  "equals_hhi_ratio_real_over_ours": true
 }
}