Statehold: A Reference Architecture for Governed, Model-Independent AI-Assisted Project Work

Statehold is a reference architecture for doing project work with AI models without letting any model or platform own the project's state. The state of a project, meaning its current status, valid decisions, rules and working files, lives in a plain-text file store controlled by the human owner. Models are configured for stateless work where platform controls permit: platform memory, cross-chat search and training on user data should be switched off, and every session rebuilds its context by following the project instruction, the short router, the current status and only the files the task needs. Statehold separates Project Truth (valid status and decisions) from Research Evidence (analyses, reviews, AI outputs). Conflicts are resolved by authority, not by file date. A missing required source leads to an explicit report instead of a reconstruction from model knowledge (fail closed). Verification is proportionate to risk, from single-model work (V0) to blind, independent solving by a second model, preferably from another model family (V2). Critical actions need explicit human approval and are blocked technically where the platform allows it. A contract version shared by platform adapters and the router blocks persistent changes when the two disagree. Storage, models and user interfaces are interchangeable, up to fully local operation with open-weight models. This document specifies the architecture, its workflows, its security considerations and an acceptance-test protocol. Empirical results are not yet part of this version.