Research & benchmarks

Alongside Thoth, the open work includes environments for evaluating agents, model architecture experiments, and tools for recording training and inference runs. These projects are maintained by HumanDotPy.

Explore the code

• FactorioAgentBench — a persistent Factorio benchmark and agent evaluation platform.
• ArchitectureResearchBench — an environment for model architecture research.
• RolloutPlane — a control plane and evidence ledger for agentic reinforcement learning.
• GradientFoundry — an evidence platform for pretraining and post-training architecture research.