← VRAXION
FROZEN
B

Block B · L1 Merger

Byte-Pair Merger

A single-matrix mirror autoencoder that takes two L0 latents and merges them into one 32-dim representation. 100% lossless across all 65,536 byte pairs.

FROZEN · v5.0.0-β.2
↓

Block B · Architecture

Single-W mirror design

L0-A out
16
+
L0-B out
16
→
Merged latent
32
→
Mirror back
32

One shared weight matrix W (32×81). Encoder and decoder are mirror-tied through this single W — the smallest possible parameterisation for lossless merge.

Two L0 outputs (each 16-dim) are concatenated to form a 32-dim input. The merger's single W matrix projects this into 81 hidden neurons with C19 activation, then mirrors back to reconstruct both L0 vectors. The breakthrough finding (2026-04-19): this H=81 single-W architecture tied to 100% lossless with only 2,592 weight cells — half the cell count of the previous absorb_float champion.

Shape
32 → 81 → 32
Weight matrix
1 × (32×81)
Weight cells
2,592
Activation
C19
Input
2 × 16-dim L0
Output
32-dim merged

Block B · Results

Compression progression

100%
Lossless on all 65,536 byte pairs
65,536
Pairs tested
3,440 B
Huffman-packed size
2,592
Weight cells
2,422 B
Shannon floor

Size progression to packed binary

fp32 original
11.20 KB
fp16 halved
5.60 KB
Huffman-packed
3.36 KB
Shannon floor
2,422 B

42% gap remains between current Huffman pack and the theoretical Shannon floor.

Block B · Artifacts

Deploy & reproduce

Packed binary artifact (current champion)
output/merger_single_w_huffman_pack/packed_model.bin
3,440 B Huffman-packed · 100% lossless
Native 7-bit alternative (Cluster 18, 2026-04-19)
identity H=120 · 7-bit W · 3-bit b1/b2 · fp32 α → full variant page
3,421 B native (no decode step) · 100% lossless · pending deploy pipeline

Reproduce — two steps

python tools/diag_byte_single_w_huffman_pack.py
python tools/diag_byte_huffman_independent_verify.py
Packed size
3,440 B
Shannon floor
2,422 B
Gap to floor
~42%
Version
v5.0.0-β.2

Block B · Roadmap

What comes next

→
Close the 42% gap to the Shannon floor: investigate arithmetic coding or range coding as a drop-in replacement for Huffman.
→
Extend merger depth to L2 (byte-quadruples) using the same single-W mirror pattern to continue the hierarchical compression chain.
→
Cluster 13 wiki entry documents the Cluster 10 "73% ceiling" overturning — the ceiling was a single-seed artifact, not a fundamental limit.

Related clusters

Cluster 13
← A Byte Unit All Blocks C Tokenizer →