Answers typed questions about a grid -- "which way should I move?", "is north
clear?", "how far is the goal?" -- with calibrated probabilities and no token
decoding. A from-scratch reimplementation of the idea behind NanoJev, rebuilt
around the thing that made the original fail at Maze: planning depth. The
backbone is a min-plus recurrence whose fixed point is a shortest-path
distance by construction; the three Jev primitives (Choice, Score, Boolean)
are kept as the interface.
Trained for 9,000 steps on one Apple M4 Max. On the held-out test split it
picks the BFS-optimal maze move 141 times out of 141 and the oracle snake move
89 out of 92, and the planner's distance field matches BFS to within 0.01
cells after a single global rescale.
This repository is the release: the weights (balanced.pt and best.pt, through
Git LFS), the model code, the training and evaluation scripts, the frozen
splits every number is measured on, and a model card whose tables are
generated from the JSON besid…
