Record: 0.8335 BPB — DualHash + AdaMuon + MoE + SDClip (3-seed mean)

[Karen042009]

New submission for the 10min/16MB track.

Key features:

• DualTokenHashSkip: Bigram skip connections with dual hash tables.
• LayerScale Recurrence: Recurrent structure with learnable LayerScale coefficients.
• SharedMoE: Hybrid MoE with shared and specialized experts.
• AdaMuon Optimizer: RMS pre-conditioning and Riemannian Newton-Schulz orthogonalization.
• Dynamic MSE SDClip: Optimal INT6 quantization search.
• Score-First TTT: Compliant 2-pass Test-Time Training.

Detailed results and architectural information are available in the included README.md.