MineChoice fresh-world completed-house checkpoint 165
This model has not beaten Minecraft. No held-out full-game success rate or server-verified Ender Dragon kill exists for this checkpoint.
This immutable 73-input, 45-action categorical MLP was trained locally on real Minecraft Java 1.20.4 survival rewards. It reads structured game state, not pixels, and has no teacher fallback. Coded skills execute its selected actions; the weights alone are not a standalone game bot.
Episode 20260928-132431-train-1881560440 began on natural seed 1881560440 with empty
inventory and no privileged gameplay setup. In that fresh-world episode the
bot finished one 52-part house at {'x': -33, 'y': 65, 'z': 13}. The reward transition
records completed houses rising from zero to one; the saved world also records
the completed build. The episode stopped at that house milestone, alive, with
85 real reward transitions and total reward
26.9760. They trained checkpoint 165, with
7743 cumulative optimized transitions. This is a training
milestone, not a measured repeat rate or proof this checkpoint will build a
house again on a different seed.
model.safetensors and policy.json contain identical weights. config.json
gives the feature and action order; milestone_evidence.json records the
scope of the claim. Bundled policy files match the episode's source snapshot.
Training used a local RTX 3090, and inference does not require a hosted API.
No world save, private credential, or raw gameplay log is included.
An iron pickaxe, diamond acquisition, Nether entry, End entry, and game completion were not demonstrated in this episode.