MineChoice visible item recovery 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 local reward-training checkpoint 58 from Minecraft Java 1.20.4 survival play. The 73-input, 45-action MLP uses direct structured state, not pixels. Coded skills execute categorical decisions, with no teacher fallback at inference. The classifier is not a standalone game bot.
Episode 20260928-025947-train-persistent-001 continued a naturally generated world with server-preserved
inventory. The bot returned to an earlier death site, gained 10 inventory items
in one recovery action and 2 in the next, and then observed zero remaining
visible drops at the site. The exact identities of all original lost items were
not verified. This is a training trace, not a measured recovery success rate
or proof this checkpoint will repeat it. The same episode later encountered an
unreachable crafting-table failure; it was interrupted, and its 89 recorded
reward transitions trained this checkpoint. milestone_evidence.json records
the scoped evidence.
model.safetensors and policy.json contain the same neural weights.
config.json records input and action order. The bundled policy schema is the
episode's source snapshot. Training ran locally on an RTX 3090; inference does
not require a hosted service.
This checkpoint has not demonstrated an iron pickaxe, diamond acquisition, Nether entry, End entry, or game completion. No world save, credential, or raw gameplay log is included.