GSA + QK-norm
This run used 24 layers with 64 slots, per-head QK RMSNorm scaled by 1.2 before SiLU, and a chunk-128 FP64 recurrence. It trained for 46,875 updates (49.152B tokens) at seed 42. Final periodic validation BPB was 0.685242, independent final validation BPB was 0.682638, and CORE ×100 was 28.1720. Its cumulative charge of 4.689438 GFLOPs per input token is 0.36468% above the SSSL anchor.
The submission includes all 189 validation points, six CORE-22 evaluations with per-task scores, and the contributor’s cumulative audit. The audit identifies the frozen source and metrics archive by hash. The submitted record contains no segment-by-segment resume log or trainable implementation for this variant.