Experimental work: 12L GQA MIXED-BIT TOWER (x5): 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. COMPRESSION PACKAGING EXPERIMENT (x1): architecture family detected in measured rows; exact flags are preserved in the row details. TRACE AND FLOW-CAPTURE WRAPPER (x1): architecture family detected in measured rows; exact flags are preserved in the row details.
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. 11L GQA U-NET BODY (x1): 11-layer dim-512 GQA U-Net transformer base with XSA, BigramHash, value embedding, and SmearGate. Specs: num_layers=11; model_dim=512; vocab_size=1024; num_heads=8; num_kv_heads=4; head_dim=64.
Day 021 measured 10 architecture/test rows against VAL_BPB, with 9 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; SmearGate.
No new personal best. Best valid of the day = 1.1057 (train_seed444.log); best-before-today = 1.0672 still stands. 1 crash-class outlier (4.21) excluded.