Summary

• 11L 512d MLP3x with int6+zstd-22, SmearGate, BigramHash, OrthoInit, Muon WD=0.02, EMA(0.997)
• FP16 tied embeddings, seq2048, sliding window eval stride=256
• Mean val_bpb: 1.1497 ± 0.0004 (3 seeds: 1337, 42, 7)
• Artifact: ~14.8MB

Test plan

• 3-seed validation on 8xH100 SXM (all under 16MB, all under 600s)
• train.log included (seed 7, val_bpb=1.1495)
• Extensive ablation documented in README (AttnRes, depth recurrence, seq-len curriculum, TTT)