Matt Rubens’s jevcraft uses Jev to choose Minecraft actions from menus built by ordinary code. Its detailed failure reports show how much depends on the options, their consequences and memory of earlier attempts.

Give the model consequences it can compare

  • Code supplies feasible actions, health, threats, equipment and estimates of damage, time and progress toward the goal.
  • A documented fireball death exposed a missing cost: being pushed off a ledge into lava.
  • Cover once looked attractive because its description priced damage while omitting that it earned no blaze rods, an item needed to advance.
  • The maintainer changes the world description and action menu; Jev itself stays unchanged.

These are the maintainer’s diagnoses of recorded failures, not controlled measurements of each fix’s effect.

Remember failure across the whole bot

A shared ledger records what was tried, where and what changed. Failed routes can rest temporarily, and exhausted questions send their failure to a higher-level planning question.

  • Waiting also needs an outcome. Staying alive in a sealed pocket can still mean making no progress.
  • Branch questions travel together in one request; code follows the selected path afterward.
  • The tree helper rejects unavailable choices and marks changed-state responses stale. The wider controller has a narrow exception for continuing an action already underway.

“None good” still becomes an action

Gameplay questions can offer none_good to expose missing moves. Selecting it logs the gap, then host code chooses a listed action anyway.

The current implementation goes beyond taking the next-highest probability. It can avoid options priced as lethal, skip recent failures or preserve the current activity. That policy deserves the same review as the model’s answer. Typed output is not a correct decision.

Read the reported run separately from the tests

The README’s September 28 account reports 15 of 40 fresh worlds reaching the Nether, with no completed game. It reports 186 ms median answers and about 1.7 cents per bot-hour in Jev input charges. Those gameplay, timing and billing results were not reproduced here; bot-hour pricing excludes game-server compute and development work.

Verification: seven repository client tests and four additional offline tree checks passed at 1eecd08f88fa2d41bd35f6906746e050fea4c2d0. Calls used test doubles; no Minecraft trial, live Jev inference or security audit was performed.