# rob-rbyte-v1

Residue router for the SAIR Modular Arithmetic Challenge. Entry class
`model.ResidueRouterV1`, output base 256.

The shipped specialist covers every prime p <= 251 (tiers 1-2). Operands are
reduced mod p inside `predict_digits` (the two-argument normalization used by
both reference models); a ~2.9M-parameter network then classifies the product
residue, emitted as one base-256 digit. Inputs with p > 251 return [0].

Training data is the complete input space for these tiers: all 995,777
triples (x, y, p) with p prime <= 251 and x, y in [0, p). Accuracy on the
public benchmark and on fresh-seed local runs: tier 1 = 1.000, tier 2 = 1.000.
See `manifest.json` for the full model and training descriptions.

Files: `model.py` (architecture + routing), `weights.safetensors`,
`config.json` (per-specialist hyperparameters), `manifest.json`.