Spaces:
Running
Python Wrapper Project that turns vLLM into a governed agent.
The Python wrapper called The Vessel started humbly and naively as a set of prompts; this then developed into mapping the prompts to a Three Persona Architecture (Executor / Whistelblower / Proxy); then pathways were explored to give it teeth and turn a capable inference engine like vLLM into a governed agent.
The capable inference engine (in this case vLLM) is still doing the heavy lifting of PagedAttention and token generation, but it is no longer "free-wheeling." It is being "interleaved"—the Vessel pauses the engine at structural checkpoints, evaluates the logic, corrects it if necessary, and then tells the engine to continue.
The wrapper applies a "Skin" (EBNF Grammar) and "Rules" (PDDL logic) over the model so that every single word is filtered through your Governance logic before it comes back to the user.
If anyone is interested, let me know. The project is on Codeberg and Github. Thanks.
My aim is to make us less vulnerable to ourselves (less vulnerable to LLMs).