{
  "entry_class": "model.ScanRegisterMachine",
  "output_base": 32,
  "framework": "pytorch",
  "model_description": "Scan-Register Machine, 68k parameters total. Two learned scan units over base-32 limb registers: (1) an adder whose carry propagation is a learned 3-element monoid operator run inside a Blelloch prefix-scan combinator, and (2) a conditional-subtract unit with learned comparison (MSB-first) and borrow monoids. Units are composed by a fixed bit-streaming double-and-add schedule (register width = limbs(p)+1, raw operands streamed as bits), with argmax re-discretization between steps. At load/predict time, the shipped neural weights are evaluated over the complete finite primitive domains (15,387 encoder/operator/resolver cells) to materialize learned class transition and resolver tables; inference scans only over those learned class IDs. Both the comparison and borrow paths are produced by the trained reducer weights; Python supplies no mapping between their semantic classes. The complete extracted tables have 0 mismatches against their finite specifications. Inputs are independently converted to base-32 limbs/bits in the per-argument preprocess hooks; no operand reduction occurs there, and the emitted digits come directly from tables extracted from the learned weights.",
  "training_description": "Reproducible checkpoint trained from random initialization (seed 20260810) by the included train.py (also mirrored as scripts/retrain_finite_cells.py in the source repository). Random base-32 limb embeddings were frozen; the encoder, learned 3-state monoid operator, and digit resolver MLPs were optimized on their complete finite primitive domains using MSE to a fixed 3-vector codebook for state-valued components and cross-entropy for digit outputs. The training labels specify carry, comparison, borrow, composition, addition-digit, and conditional-subtraction-digit behavior at the local cell level; no end-to-end modular answers or evaluation cases are used. Training stops only when all extracted tables are exact. This artifact was audited at 15,387/15,387 cells and randomizing the weights collapses accuracy. The script regenerates a fresh exact checkpoint in hundreds of optimizer steps."
}