{
  "artifact": "Blackfrost-AI/CYBER-FROST-3.8-FP4-Local",
  "clean_model_name": "CYBER-FROST-3.8-FP4-Local",
  "release_date": "2026-10-17",
  "release_type": "quantization re-export",
  "training_added": false,
  "lora_added": false,
  "new_fine_tuning": false,

  "lineage": {
    "foundational_checkpoint": {
      "repo": "Qwen/Qwen3.8-Flash-Next",
      "revision": "de4b8e4d43b917e7706784d8bb445c9af86a3540"
    },
    "security_adaptation": {
      "stage": "BLACKFROST-3.8-FLASH-BF16",
      "description": "security-domain fine-tuning followed by full BF16 merge"
    },
    "behavioral_stage": {
      "description": "Blackfrost-AI behavioral derivative targeting reduced false-refusal friction in authorized security workflows",
      "process_distributed": false
    },
    "bf16_source": {
      "repo": "Blackfrost-AI/CYBER-FROST-3.8-BF16",
      "revision": "5321904427c4ef54df8a667edcbc2d1184e4286e"
    },
    "nvfp4_v2_export": {
      "repo": "Blackfrost-AI/CYBER-FROST-3.8-NVFP4-V2",
      "date": "2026-10-02",
      "description": "routed expert projections to ModelOpt NVFP4 W4A4 g16 with per-expert shared gate/up global scales"
    },
    "fp4_local_export": {
      "date": "2026-10-17",
      "description": "this repository"
    }
  },

  "conversion": {
    "format_reference": {
      "repo": "nvidia/Qwen3.8-Flash-Next-FP4-Local",
      "role": "on-disk format and index-structure reference only; not weight or behavioral lineage",
      "shards": 21
    },
    "modelopt_version_reported": "0.47.0rc0",
    "pipeline": "DWM_Models/runs/cyber-frost/20261017-fp4-local/convert_fp4_local.py",
    "workers": 16,
    "wall_clock_minutes_approx": 9,
    "tensor_provenance": [
      {
        "class": "copy_v2",
        "description": "routed expert packed weights and FP32 outer scales",
        "count": 221184,
        "bytes_gib": 56.25,
        "method": "byte-copy from CYBER-FROST-3.8-NVFP4-V2"
      },
      {
        "class": "csf_fixed",
        "description": "per-16-row base byte plus packed 4-bit offset codes",
        "count": 73728,
        "bytes_gib": 3.6,
        "method": "fresh CSF encode of V2 scale planes"
      },
      {
        "class": "csf_exceptions",
        "description": "out-of-range scale entries as flat_pos | (exact_byte << 24)",
        "count": 73728,
        "bytes_mib": 92,
        "method": "fresh CSF encode; counts data-dependent"
      },
      {
        "class": "ple_w",
        "description": "PLE n-gram table packed four-bit weights",
        "count": 128,
        "bytes_gib": 23.8,
        "method": "fresh NVFP4 two-level encode of BF16 source"
      },
      {
        "class": "ple_s",
        "description": "PLE n-gram scale planes",
        "count": 128,
        "bytes_gib": 3.0,
        "method": "fresh two-level encode"
      },
      {
        "class": "mxfp8_w",
        "description": "attention and shared expert projections",
        "count": 438,
        "bytes_gib": 2.9,
        "method": "fresh MXFP8 encode, ModelOpt semantics"
      },
      {
        "class": "mtp_w",
        "description": "MTP W4A16 experts",
        "count": 512,
        "bytes_gib": 1.2,
        "method": "fresh encode of fused gate_up/down"
      },
      {
        "class": "preserved",
        "description": "norms, embeddings, lm_head, PLE metadata",
        "bytes_gib": 3.9,
        "method": "source bytes; norms upcast BF16 to F32 per reference dtype"
      }
    ]
  },

  "corpus_provenance": {
    "note": "Inherited from the BF16 parent; the FP4-Local conversion added no new fine-tuning data",
    "teacher_policy": "frontier-scale only; teachers below the 753B-parameter class excluded",
    "teacher_manifest_complete": false,
    "independently_bound_subset": "one security subset bound to a Qwen3.8 2.4T teacher",
    "licensing_review_status": "in progress for mixed-source corpus"
  },

  "known_deltas_vs_reference": [
    "nvfp4_csf_exceptions tensor counts differ from NVIDIA's Qwen release because the encoded planes differ (different weights). Storage format, codec, and reconstruction semantics are identical.",
    "Residual 0.4-0.6% adjacent-step differences in freshly encoded packed weights (PLE, MTP, MXFP8 paths) stem from QAD rounding edge cases on different weight distributions. The recipe is byte-validated against NVIDIA's own outputs.",
    "quantization_config.quantized_layers ships verbatim from the reference config and includes vestigial mtp.layers.48.* keys present in neither model's tensors."
  ]
}