MineChoice jungle-world iron-pickaxe checkpoint
This model has not beaten Minecraft. There is no held-out full-game success rate or server-verified Ender Dragon kill.
This immutable local reward-training checkpoint 86 uses a 73-input, 45-action categorical MLP on direct structured Minecraft Java 1.20.4 state, without pixels or a teacher fallback at inference. Coded skills execute its decisions; the classifier is not a standalone game bot.
Seed 2026092801 first started as a natural survival world with an
empty inventory and no privileged setup in episode 20260928-041735-train-2026092801. The
verified iron pickaxe was crafted later, in continued-world episode
20260928-044339-train-persistent-001, with server-preserved inventory after intervening training and
code repairs. It was not one uninterrupted fresh-world completion trial.
That episode logged 72 real reward transitions and a +29.36
episode reward before stopping at its configured iron-pickaxe checkpoint.
This is one training milestone, not a measured repetition rate.
model.safetensors and policy.json contain the same neural weights.
config.json records the exact input and action order, and
milestone_evidence.json records the verified scope. The bundled policy schema
is the milestone episode's source snapshot. Training ran locally on an RTX 3090.
No world save, credential, or raw gameplay log is included.
Diamond acquisition, Nether entry, End entry, and game completion remain unverified for this checkpoint.