Experimental work: 12L GQA MIXED-BIT TOWER (x1): 12-layer GQA tower: 8/4 GQA, RoPE-16, Bigram 2048, XSA on the last 11 layers, and mixed-bit packaging. Specs: num_layers=12; num_heads=8; num_kv_heads=4; rope_dims=16; eval_seq_len=2048; train_seq_len=2048. CRAWLER-SAFE TRAIN/EVAL SURFACE (x1): Crawler-safe training/eval surface around a widened shared-crawler stack with final int6 and int8 checks. Specs: model_dim=512; vocab_size=1024; num_heads=8; num_kv_heads=4; BigramHash; XSA.
Neural work: 12L GQA MIXED-BIT TOWER (x2): 12-layer GQA tower: 8/4 GQA, RoPE-16, Bigram 2048, XSA on the last 11 layers, and mixed-bit packaging. Specs: num_layers=12; num_heads=8; num_kv_heads=4; rope_dims=16; eval_seq_len=2048; train_seq_len=2048.
Day 028 measured 4 architecture/test rows against VAL_BPB, with 4 valid rows and 0 TTT leakage readings excluded. Techniques to copy/reference: attention int5; mlp int6; aux int6; embed int8; other int8; Brotli artifact; 16 MB artifact cap; BigramHash.
No new personal best. Best valid of the day = 1.0900 (Fix bank quantization and quant-escape evaluation bug); best-before-today = 1.0672 still stands.