DishwasherBench is a dishwasher-loading game built on a physical simulation of one real machine, a Bosch SMV4HAX40G. It was created to benchmark language models and can also be played by people in a browser. The game inside is called Spotless, and its packages are named @spotless/*.
A player loads a 3D model of the dishwasher using rigid-body physics. A wash on its Eco 50 programme is then simulated, covering jets, droplets, run-off, soil and detergent chemistry. The player earns points for every item that comes out clean. From the second level onward, not everything fits, so the player must choose what to load and how.
This is an early release, version 0.1.0. The simulation is calibrated on a single test wash of the real machine. A score shows whether a player packed well and placed cups, glasses and pots mouth down. It does not yet show whether plates and cutlery were stacked so that they wash clean. Loading physics and replayable episode records are described as solid, while the scoring of plates and cutlery is described as rough.
The project is built and maintained by Composo and is published on GitHub. It includes a course of seven levels ordered by difficulty, three ways to play via the web page, an MCP server for agents, and a benchmarking harness. Two models have played six episodes so far at the official tier, and no human baseline has yet been recorded.
Source: Hacker News · Summarized by HeadlinesBriefing