The VSKI Whitepaper: Auditable Intelligent Systems on a Stock-and-Flow Substrate

Abstract

VSKI (the Vector Semantic Knowledge Index) is a self-hosted platform for building data-driven intelligent systems, and VSKI Lab is its modeling environment. Together they implement one idea: **the symbolic layer holds the truth and asserts it, while the neural layer proposes and is verified**. The core formal object is a directed graph of stocks and flows in which conservation of quantities is asserted on every tick, so that accounting errors, probability mass that appears from nowhere, or money that vanishes are rejected by construction rather than discovered after the fact. The platform provides durable storage (typed collections, SQL views, full-text and vector search), event-sourced workflows, and a records API; the Lab provides the modeling loop: compose the system, run it tick by tick, fence and audit agent interventions, capture trajectories, declare objectives, and train replacements inside the same harness. This whitepaper states the design principles, formalizes the model, and demonstrates the framework on two use cases: an academic one (a probability space as a flow network, where flow conservation *is* the law of total probability) and a business one (a delivery kitchen, where conservation *is* double-entry bookkeeping and the interesting behaviour is queueing).