{
  "claim": "The shipped needle3.cact stores 115 tensors at CQ2 and 7 at CQ4, and its tensor directory reconciles to 35,335,380 bytes exactly. The locally installed cactus-needle package cannot write CQ2 at all \u2014 export.py raises on any bits other than WEIGHT_BITS, which is 4 \u2014 so a 20-layer archive built by `needle build --lora` is 63,437,076 bytes: 1.80x the shipped file, and with the confidence head deleted. Every row is arithmetic over the archive's own published directory records, cross-checked by reconstructing the shipped total to the byte.",
  "method": "One HTTP range request for the first 128 KiB of needle3.cact reads the 196-byte header, the 28-float codebook and all 581 tensor records (dtype, shape, offset, nbytes, group, bits). Blob sizes are recomputed from the documented CQ layout (out * in_pad * bits / 8 packed indices, plus out * in_pad/128 FP16 group norms, in_pad = ceil(in/128)*128), 64-byte aligned, and summed. Reconstructing the shipped scheme reproduces 35,335,380 bytes with zero error, which is what validates the W4 figures.",
  "source": "https://huggingface.co/Cactus-Compute/needle3 + cactus-needle 3.0.4 (needle/model/export.py, needle/model/quantize.py)",
  "captured": "2026-09-22",
  "note": "Depth rows assume the documented ladder order (0, 19, 9, 14, 4, 6, 11, 16, ...) and drop engram sites whose host layer is not kept. The 20-layer platform row is measured; the smaller platform rows are computed, because Cactus publishes no pre-sliced archives.",
  "columns": [
    {
      "key": "build",
      "label": "build"
    },
    {
      "key": "depth",
      "label": "layers",
      "align": "right"
    },
    {
      "key": "bits",
      "label": "weight bits"
    },
    {
      "key": "head",
      "label": "confidence head"
    },
    {
      "key": "bytes",
      "label": "bytes",
      "align": "right"
    },
    {
      "key": "mb",
      "label": "MB (10^6)",
      "align": "right"
    }
  ],
  "rows": [
    {
      "build": "shipped needle3.cact (reconstructed, 0 byte error)",
      "depth": 20,
      "bits": "embedding=4, mhc=4, head=4, default=2",
      "head": "present, 37,276 B",
      "bytes": 35335380,
      "mb": 35.34
    },
    {
      "build": "needle build --lora (local)",
      "depth": 20,
      "bits": "4 everywhere",
      "head": "deleted",
      "bytes": 63437076,
      "mb": 63.44
    },
    {
      "build": "needle build --lora --layers 16",
      "depth": 16,
      "bits": "4 everywhere",
      "head": "deleted",
      "bytes": 51422676,
      "mb": 51.42
    },
    {
      "build": "needle build --lora --layers 8",
      "depth": 8,
      "bits": "4 everywhere",
      "head": "deleted",
      "bytes": 27393876,
      "mb": 27.39
    },
    {
      "build": "needle build --lora --layers 4",
      "depth": 4,
      "bits": "4 everywhere",
      "head": "deleted",
      "bytes": 15379476,
      "mb": 15.38
    },
    {
      "build": "needle build --lora --layers 2",
      "depth": 2,
      "bits": "4 everywhere",
      "head": "deleted",
      "bytes": 13329108,
      "mb": 13.33
    },
    {
      "build": "platform scheme at 16 layers (computed)",
      "depth": 16,
      "bits": "embedding=4, mhc=4, head=4, default=2",
      "head": "retrained",
      "bytes": 28948884,
      "mb": 28.95
    },
    {
      "build": "platform scheme at 8 layers (computed)",
      "depth": 8,
      "bits": "embedding=4, mhc=4, head=4, default=2",
      "head": "retrained",
      "bytes": 16175892,
      "mb": 16.18
    },
    {
      "build": "platform scheme at 2 layers (computed)",
      "depth": 2,
      "bits": "embedding=4, mhc=4, head=4, default=2",
      "head": "retrained",
      "bytes": 8636052,
      "mb": 8.64
    }
  ]
}