manifest.json

{
  "entry_class": "model.FastModularModel",
  "output_base": 2,
  "framework": "pytorch",
  "model_description": "A 260,929-parameter modulus-conditioned recurrent model with a 64-dimensional input projection and a two-layer, 96-hidden-unit bidirectional GRU. After pairwise a mod p and b mod p normalization, the learned cell repeatedly predicts the radix-2 Horner transition over binary state, multiplicand, and modulus channels, then emits the final state as base-2 digits.",
  "training_description": "Trained independently from random initialization using AdamW on synthetic exact transitions from 8 through 2,048 bits. The curriculum mixed random moduli, small primes, wrap boundaries, on-policy Horner states, sparse powers, all-one prefixes, and width-boundary trajectories, with held-out exact-match and signed-margin checkpoint selection. The final checkpoint is a low-rate continuation hardened on anchor-biased structured trajectories. No pretrained, external, hand-initialized, or competitor weights were used."
}