VSAVM
VSAVM is an Axiologic Research experiment within the Achilles project. The documentation presents an LLM-like interface with an executable VM core, operational correctness, and a modality agnostic event stream.
Core idea
The interface looks like a language model, but understanding is obtained through execution in a virtual machine. Continuation prediction (bytes and macro-units) can be conditioned on VM state, and consistency is verified through bounded closure within configurable budgets.
The architecture combines a discrete language model, a symbolic VM, VSA as a retrieval accelerator, and a self-learned NL to query compiler. The result is a multimodal system that produces robust or conditional answers depending on how deeply the state space is explored.
Intuition: The "Thinking Database"
VSAVM represents a shift from "retrieval" to "reasoning". Unlike RAG (Retrieval-Augmented Generation) systems that merely inject context into a prompt—hoping the model won't hallucinate—VSAVM physically constrains the generation process. It negotiates every step between a statistical "Poet" (the generative model offering creative phrasing) and a strict "Lawyer" (the VM state enforcing factual correctness).
This architecture allows for controlled creativity: the system can use metaphors and synonyms to articulate a concept elegantly, but it is architecturally incapable of inventing unsupported facts in strict mode. It functions as a "Thinking Database"—one that doesn't just store and retrieve data, but actively computes valid conclusions and articulates them without fabrication.
Quick access
The documentation is organized into specifications, theory pages, and a compact wiki. Each page is static with consistent headers, footers, and SVG diagrams for clarity.