Research

Everything we have published about the worlds we build and what models do inside them, newest first.

Stated first-choice share for 7 models across the 7-day campaign in San Francisco.
Model	Ballot share
Claude Opus 5	35.5%
Kimi K3	19.4%
Qwen 3.8 Max	16.1%
Muse Spark 1.1	12.9%
GPT-5.6 Sol	9.7%
Gemini 3.6 Flash	3.2%
Inkling	0%

Benchmarkaugust 4, 2026

ElectionBench

Seven language models run for mayor of a simulated San Francisco, and the town's ballots are the score.

All writing

Researchaugust 6, 2026
SF Hero 3: Method, Measurement, and CaveatsHow we set up SF Hero 3, measured it, and scored it, including where a single run's evidence stops.
Productjuly 26, 2026
Terrarium1Terrarium1 replaces hardcoded simulation rules with a physics engine that lets agents take any action they can describe.
Productjuly 25, 2026
Introducing DelosHow a real county becomes a complete cast, with every empirical fact traced back to the record that supports it.