Products
Open Source Research Tools
We build tools that we use in our own research, and release them for anyone to use.
Constraint Generation
The constraint stack
A typed engine, and the system-agnostic interface built on top of it.
Built on AUFBAU
Proposition 7
System-agnostic, formally constrained generation for LLMs using AUFBAU. Type-aware constrained decoding, a benchmark runner, and built-in typed grammars for STLC, ML, and imperative fragments. Named for the last line of the Tractatus.
“Wovon man nicht sprechen kann, darüber muß man schweigen.
L. Wittgenstein · Tractatus Logico-Philosophicus, 7
Tooling
Research Software
Open source tools built for and alongside our research.
SLOP
Infrastructure for low-budget comparative study of AI slop and colonial imagery. Custom difusion models implementations for field probing and latent space exploration.
SLUT
Static Linear Untiful Tensors. A Rust crate for working with vectors and matrices with physical dimensions. Basically uom but in a vector space. Zero-runtime code, fully compile-time.
HRML
The markup language and framework used to build this website. A custom templating engine with component hoisting, MDX support, and static site generation.
Political Alignment
Tools to study LLM political alignment. Automated political bias evaluation on a political compass, and embedding space analysis to measure political distances between concepts.
Open Source
Committed to Open Research
We believe research tools should be free, accessible, and reproducible.
Open by Default
Every project we release is open source under MIT. Our research code, tools, and benchmarks are available for anyone to inspect, fork, and build upon.
Community Driven
We welcome contributions, issues, and discussions. Our repositories are open for collaboration with researchers, engineers, and anyone curious about formal methods.
Reproducible Science
Benchmarks come with configs, Docker artifacts, and detailed reproduction instructions. If we published a result, you should be able to verify it.
Get Involved
Explore our work, open an issue, or start a conversation.