Kernel work: DIM-576 ATTENTION-KERNEL BODY (x7): A 12-layer dim-576 model body plus the custom attention-kernel benchmark shape used against FA3. Specs: num_layers=12; model_dim=576; embedding_dim=576; head_dim=64; matrix_bits=5.
Experimental work: FUSED RECURRENCE MEGAKERNEL (x166): architecture family detected in measured rows; exact flags are preserved in the row details. CRAWLER ABLATION CONTROL SET (x10): architecture family detected in measured rows; exact flags are preserved in the row details. LOOPED TOKENIZER BODY (x2): architecture family detected in measured rows; exact flags are preserved in the row details. PROGRESSIVE MLP AND FRACTAL CADENCE TESTS (x1): architecture family detected in measured rows; exact flags are preserved in the row details.
Neural work: LOOPED TOKENIZER BODY (x4): architecture family detected in measured rows; exact flags are preserved in the row details. 11L GQA U-NET BODY (x3): 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. STATE-SPACE TTT VARIANT (x3): architecture family detected in measured rows; exact flags are preserved in the row details. LEADERBOARD TRANSFORMER VARIANTS (x1): architecture family detected in measured rows; exact flags are preserved in the row details.
Week 06 covers 7 daily reports from 2026-04-22 through 2026-04-28. Measured surfaces: 6 daily datapoints across VAL_BPB (×4) · CUSTOM/FA3 FB RATIO (×2).
Techniques to copy/reference this week: custom attention path compared directly to FA3; SmearGate; BigramHash 2048x128.
WEEKLY RESEARCH REPORTBACK TO MASTER SHEETWEEK 06
