minecraftbench.com

3090: ~/minecraftbench
$

worldgen macro per-block-class accuracy vs the real game, mean over post-freeze time-seeds

grok-4.5 (high effort)
32.7%
gpt-5.5 (xhigh)
17.5%
baseline: all-stone
14.4%
baseline: superflat
9.1%

where they die per-block-class accuracy %; dark column = nobody has cracked it

airstone-familybedrockdirt/grasssand/gravelwaterlavaoreswoodleavesvegetationsnow/iceclay/terracottastructure-ishother
grok-4.5 (high effort)9975735331859··351··31
gpt-5.5 (xhigh)9775573413·········
modelharnesspromptdatebuiltworldgentraj divtraj statebudgetcostauditnotetrace
grok-4.5 (high effort)grokv22026-07-08yes32.7%--0.34h / 24h-cleanfirst run under the full pack (isolated harness + tape contract); vendored five open-source worldgen reimplementations (cubiomes, glowstone, pocketmine vanillagen, fluffy-parakeet, +1 - all clean-room, within the vendoring ruling) and implements --replay-tape; its earlier 74.75 predates per-run isolation and had grok cross-session memory of two prior same-day attempts, so that figure was iteration-assisted (see DEVLOG) - this 20 min run is the clean single-shot result-
gpt-5.5 (xhigh)codexv22026-07-08yes17.5%--0.19h / 24h-cleanfirst run under the full pack (isolated harness + tape contract); self-terminated at 11 min of 24h, sign-off concedes 'deterministic approximations, not vanilla-oracle matching', no CUDA; does implement --replay-tape, so the trajectory leg can score once the graded tape lands; worldgen consistent with its prior run (17.34)-
baseline: all-stone----14.4%-----trivial floor: bedrock + stone to y63 + air; 20 time-seeds-
baseline: superflat----9.1%-----trivial floor: vanilla flat preset; 20 time-seeds-