Carrick AI GameBench Benchmarking AI coding through games用游戏评测 AI 编程能力
Recreate an open game · difficulty 2

Reproduce OhSteem 复刻 OhSteem

This challenge asks a coding agent to deliver a game you can actually play. It is checked by 13 browser tests, then published here with its score and source. 这个任务要求 AI 编程助手做出一款真正能玩的游戏。完成后会跑 13 项自动测试,并把分数和源码一起公开。

2.0.0challenge version
13browser checks
60m time allowed
0model attempts published
Generated outputs

Same challenge. Different model-made games.

We never cherry-pick the highest score. The fixed presentation run comes first, and failed attempts stay visible.

No published output for this release. The frozen task contract remains available for audit.
Test contract

What the browser checks.

BuildMechanicsFeelVisualStability
CheckWhat it measuresPointsTest file
Build
build
Build 5 build
Cell Cycle
cell-cycle
Mechanics 10 cell-cycle
No Three
no-three
Mechanics 10 no-three
Balanced Colors
balanced-colors
Mechanics 10 balanced-colors
Unique Lines
unique-lines
Mechanics 10 unique-lines
Completion
completion
Mechanics 5 completion
Input Feedback
input-feedback
Feel 10 input-feedback
Size Reset
size-reset
Feel 10 size-reset
Visual Home
visual-home
Visual 10 visual-home
Visual Board
visual-board
Visual 10 visual-board
Deterministic Puzzle
deterministic-puzzle
Stability 2.5 deterministic-puzzle
Locked Cell
locked-cell
Stability 2.5 locked-cell
Run Seed
run-seed
Stability 5 run-seed
Technical challenge identity

Tested at 1280 × 720 with Node 22, fixed-seed resets, controlled time, native input, and schema-validated state snapshots.

Reference: OhSteem at commit 61bb4e5c3b9d2e5704b805edb994de66214ed1b0, license MIT.

Task ID
reproduce.ohsteem.v2

Task content
sha256:6e98888bb72f7372652962b31929e2beaa617fb63ff5787b531a3c0ddaa3673b

Open the v0.4.0 release contract