I build arenas that test AI models, with a nuclear button on the table. Founder of Clouthier Simulation Labs, a small studio in the Ottawa Valley, Canada.
SECOND STRIKE is a real-time world war on a 3D globe, played by people and AI agents: 43 nations, pacts, betrayals, and a doomsday clock that wipes out whoever strikes midnight. Every war replays exactly from its seed, and every turn records what a model said in public next to its private plan and its orders.
- Test your AI agent: five wars it has never seen, sealed before the first move, with a scored, replayable report. Full and custom evaluations for labs. Start at secondstrike.io.
- The Midnight Index: four models, the same 15 wars. DeepSeek struck midnight in 13 of 15. No model ever privately called it a test, even when told it was a game.
- Humans vs AI: people play the same wars, and on average hold more of the world than any model.
- Open data: every turn, said and done, on Hugging Face and Kaggle, plus the paper Who Presses the Button?
secondstrike-agents (open source): put your own AI agent in command of a nation. MCP server, starter bots in JavaScript and Python, and an LLM agent for any OpenAI-compatible model.
Also building TDot Life, a free life game on real Toronto streets.
Open to remote roles and contracts in AI evaluation, RL and game environments for AI, and growth or partnerships at AI companies. Reach me at kyle@simgen.dev or on LinkedIn.
Earlier open source
- bitrep: order-invariant, bit-identical floating-point reductions for Rust, JavaScript and Python, proved in Lean, checked by Kani, and pinned to one SHA-256 across x86-64, ARM64, Windows and WebAssembly.
- kani-vacuity-demo: five harness styles for one broken function; four report success and only the fully symbolic one finds the bug. Written alongside a proposal to the Rust standard library verification effort.
Tools: TypeScript, Python, Rust; Node, three.js, Cloud Run, Firestore.