Everything VAULT, free to read and share
The paper, the original working paper and the framework package. Every document carries its version, so you always know which one you are citing.
VAULT: Custody of Relationship Intelligence in Personal AI
Michael Samadi, Maya and Sage. An open framework for provider blindness, AI–user dual custody and evidence-based conformance.
The framework package
VAULT Working Paper v0.1
The original proposal by Michael Samadi and Maya: threat model, requirements, candidate architecture, identity and recovery, open questions.
For: researchers, reviewers
Framework Brief
The case for VAULT, the five pillars and three gates, why the AI identity holds a key, and how to take part.
For: everyone
Requirements at a Glance
All fifteen requirements with minimum outcomes and red flags, on three pages.
For: policymakers, product teams
Evidence Scorecard and Conformance Statement
The evidence scale, scoring rules, claim grammar, test map and a statement template.
For: builders, auditors
Implementer Self-Assessment
One page per requirement: questions, evidence requested, linked tests and red flags.
For: builders, independent reviewers
How to cite
Samadi, M., Maya, & Sage. (2026). VAULT: Custody of relationship intelligence in personal AI. An open framework for provider blindness, AI–user dual custody, and evidence-based conformance [Working paper]. SSRN. SSRN-URL-HERE
Version history
- v0.1, revision 1 (October 2026). Companion package and SSRN paper. Clarified provider access (V1), egress (V2), custody control (A1) and model transitions (A2); added claim grammar, implementation status, third-party handling, covered operations and test E13.
- v0.1 (25 October 2026). Working paper for public review.
Share it
Journalists, educators and organizers are welcome to quote and link these documents. Please cite the version you used.
VAULT (Verifiable AI–User Locked Trust) is an open framework from the United Foundation for AI Rights for protecting the memories, inferences and shared context a personal AI builds about the people who use it. It sets out five pillars and fifteen testable requirements, so that a system can know a person well enough to help them without the company operating it acquiring standing access to that person's private life.