{
  "method": "Pearson correlation of matching public safetensors slices, HTTP byte-range reads; exploratory audit by Claude Code",
  "date": "2026-10-18",
  "n_elements_per_slice": 2000000,
  "comparisons": [
    {
      "a": "Devstral2-123B",
      "b": "Mistral-Medium3.5",
      "tensor": "L0 attention.wk",
      "r": 0.9985
    },
    {
      "a": "Devstral2-123B",
      "b": "Mistral-Medium3.5",
      "tensor": "L1 ffn.w1",
      "r": 0.9943
    },
    {
      "a": "Mistral-Large-2411",
      "b": "Devstral2-123B",
      "tensor": "L0 attention.wk",
      "r": 0.8636
    },
    {
      "a": "Mistral-Large-2411",
      "b": "Devstral2-123B",
      "tensor": "L1 ffn.w1",
      "r": 0.5575
    },
    {
      "a": "Qwen3.5-27B",
      "b": "Qwen3.6-27B",
      "tensor": "L10 mlp.up_proj",
      "r": 0.8999
    },
    {
      "a": "Qwen3.5-27B",
      "b": "Qwen3.8-27B",
      "tensor": "L10 mlp.up_proj",
      "r": 0.7528
    },
    {
      "a": "Qwen3.5-27B",
      "b": "Qwen3.8-27B",
      "tensor": "L11 attention.q_proj",
      "r": 0.8711
    }
  ],
  "negative_controls": "Adjacent-layer correlations approximately0.001(Qwen) and0.007(Mistral). These are not independent same-architecture pretrained models.",
  "interpretation": "Strong weight similarity compatible with a shared checkpoint/ancestry. No producer confirmation of exact lineage. No token count or training cost can be inferred from r.",
  "limitations": [
    "Selected small slices, not full tensors/models",
    "Common initialization also possible; correlation alone does not prove saved pretraining compute",
    "Low correlation does not rule out permutation or expansion",
    "Initial experiments used resolve/main rather than immutable revisions; rerun with pinned commits before publication"
  ],
  "source": "claude-audit/LINEAGE-CAVEATS.txt"
}