We built this because we realized that "prompt engineering" isn't enough for critical AI systems (like in finance or governance). You can't just ask an LLM nicely not to delete a database—you need a runtime guarantee.

CSL-Core is a policy language designed to bring "Policy-as-Code" to AI agents.

Instead of relying on the model's probabilistic nature, CSL enforces constraints that are:

1. Formally Verified: Policies are compiled into Z3 constraints to mathematically prove they have no logical conflicts or loopholes.

2. Deterministic: The checks happen in a separate runtime engine, independent of the LLM's context window.

3. Model Agnostic: It acts as a firewall between the LLM and your tools/API.

It's currently in Alpha (v0.2). Currently working on TLA+ specifications for the dual formal verification engine and governance architecture because we believe AI safety needs mathematical rigor.

I'd appreciate any feedback on the DSL syntax and our verification approach.