Experimental work: WIDENED CRAWLER BOTTLENECK (x8): Widened crawler bottleneck body: 5 flat blocks, 1 shared crawler block, 5 flat blocks. Specs: MODEL_DIM=512; num_heads=8; num_kv_heads=4; ROPE_DIMS=16. CRAWLER-SAFE TRAIN/EVAL SURFACE (x2): 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. 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.
Neural 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.
Day 024 measured 12 architecture/test rows against VAL_BPB, with 12 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 = 1.0742 (train_seed4_4h.log); best-before-today = 1.0672 still stands.