{
  "benchmark": "BF16 full-vocabulary fidelity diagnostic",
  "distributions": 64,
  "next_token_losses": 60,
  "vocab": 248320,
  "domains": [
    "assistant_dialogue",
    "math_reasoning",
    "python",
    "physics_engineering"
  ],
  "positions": "48 through 63 after each of four 64-token prefixes",
  "bf16_ppl_60": 2.017988158399302,
  "reference_sha256": "74566152452e5b6ce7cb22cd32a80ef5ea57f6d23ad6f872132b1968b9a4a7b9",
  "panel_sha256": "decdcfe075ac3d7961c1ad72f9754f18d90bfda16d402101de725a9d846048d0",
  "rows": {
    "ciru": {
      "mean_kl": 0.030452830391296454,
      "median_kl": 0.0027612169312611666,
      "p95_kl": 0.1545270358247857,
      "max_kl": 0.33367892204649147,
      "top1_agreement": 61,
      "ppl_60": 2.084010558412682,
      "sha256": "2d2edef8de50e5781fc6e4dcc1cf9c80daaeea9097d0ba79b4340b07d4b1d279"
    },
    "agention": {
      "mean_kl": 0.11223406250141665,
      "median_kl": 0.007866695901408137,
      "p95_kl": 0.2976523841628686,
      "max_kl": 1.9122826231950454,
      "top1_agreement": 60,
      "ppl_60": 2.383803541246223,
      "sha256": "25e5f12db31908110d0e1f02e7fab1afcd618de106da6649637b58ceaf42df9d"
    },
    "unsloth": {
      "mean_kl": 0.17457136128131387,
      "median_kl": 0.003924902626586302,
      "p95_kl": 0.5572221597760761,
      "max_kl": 3.8420741561452627,
      "top1_agreement": 61,
      "ppl_60": 2.4293137420087842,
      "sha256": "fcfbe409021ae5125d905af36ce119d4702aab576db047251a173a0ce275a0db"
    }
  },
  "limits": "Small numerical panel, not a full perplexity or task benchmark. CIRU F16 KV/FA auto; competitor Q8_0 KV/FA on. MTP off. Two independent CIRU loads produce identical raw logits. Other capture-to-CIRU differences are deliberately omitted from this BF16-only table."
}