MineChoice reward-trained survival classifier — iron-ingot training checkpoint
This model has not beaten Minecraft. No held-out full-game evaluation or server-verified Ender Dragon kill has been demonstrated.
This is immutable training checkpoint 41 from local Minecraft Java 1.20.4 survival play. The 73-input, 45-action MLP uses direct structured game state, not screenshots or pixels. Its choices are trained from observed gameplay rewards with episodic policy gradient; inference has no teacher fallback. Coded skills execute its categorical actions, and deterministic feasibility masks remove unavailable actions. The classifier alone is not a standalone game bot.
The verified training episode 20260928-001732-train-persistent-001 ran in a continued naturally generated
survival world without privileged gameplay setup. It smelted five iron
ingots from collected raw iron. The episode later ended in death by wall
suffocation. The checkpoint includes that reward update and one subsequent
death/recovery update; the milestone is training evidence, not a measured
success rate or proof this checkpoint will repeat it in a fresh world. See
milestone_evidence.json for the exact before/after inventory and reward.
model.safetensors contains the neural weights. policy.json contains the
same weights plus feature order and action schema for native Node.js CPU
inference. fullgame-policy.mjs exports features, mask, and predict;
policy.mjs supplies its original eight-action schema. config.json lists
all 73 input names and 45 output actions. The model was trained on a local
RTX 3090, but runtime inference does not need a hosted service.
This checkpoint is experimental. It has not independently demonstrated an iron pickaxe, diamond acquisition, Nether entry, End entry, or game completion. No world save, private credential, or raw training log is included.