Badtheorylabs ORG

A local-first capability layer for small open models: context, planning, tools, and verification the frontier scaffolding, minus the frontier model.

Native runtime, integrations, and reproducible packaging for BTL-3 and BTL-3 Compact

BTL-2 Coder 7B is the first BTL model built for agentic code review, and the foundation for a local terminal coding agent.

Privacy First Self Improving Agent

Echo-Skeleton Perception: stateful event-driven screen perception for text-only computer-use agents

Tool Repairing model

A controlled benchmark demonstrating that LLMs (GPT-5.4, GPT-4o mini, Gemini) score at random chance on interventional causal reasoning, while exact inference achieves 100% and human experts 97.8%.