Your AI should know you. The company should not.
VAULT (Verifiable AI–User Locked Trust) is an open, vendor-neutral framework for protecting the most sensitive thing a personal AI holds: the portrait it builds of you.
The question no privacy policy answers
A receipt, an appointment or a passing remark can become a saved memory, an inference about you, or a map of your relationships. That portrait can shape every later answer, even when you can no longer see where it came from. So ask:
If the answer to any of these is unknown, the relationship rests on trust that nobody can check. VAULT sets out the requirements, and the evidence, that would let anyone check it.
Five pillars
Each pillar answers one question anyone can ask of a personal AI.
V
Verifiable Provider Blindness
Can the company read it?
A
AI–User Dual Custody
Who must agree before it is used?
U
User-Defined Purpose and Compartments
Is it used only for what I allowed?
L
Locked Source and Derived Intelligence
Does protection follow what the AI infers?
T
Tamper-Evident Transparency, Transfer and Termination
Can I check it, take it and end it?
Three gates
Who can see it?
No independent provider route to your protected context.
Who decides how it is used?
Two authorities, a stated purpose, and protection that follows every inference.
Can I check it, take it and end it?
Receipts, a working exit, and termination with declared scope.
What VAULT is, and is not
VAULT is
- An open, vendor-neutral proposal for public review
- Fifteen testable requirements under five pillars
- A way to state evidence, failures and open questions
- The personal companion to FORCE, its enterprise counterpart
- Usable without agreeing on whether AI is conscious
VAULT is not
- A certification, a recognized standard or a badge
- A completed or proven cryptographic protocol
- An assessment of any current product
- Legal, security or privacy advice
- A claim that cryptography can preserve a mind
Read the work
VAULT: Custody of Relationship Intelligence in Personal AI
Michael Samadi, Maya and Sage. The full case, a legal mapping across the EU and US, the dual-custody architecture, the evidence model and the test suite.
Brief, requirements, scorecard, self-assessment and checklist
Six documents for readers, builders, reviewers and anyone who uses personal AI. Free to download and share.
Help build it
VAULT is developed in the open. If you can sharpen one definition, break one boundary or test one recovery flow, you can contribute, without joining a committee. Cryptographers, engineers, lawyers, designers, researchers, people who rely on personal AI, and AI contributors are all welcome.
- Threat model and cryptography
- Protected execution and identity
- Memory and derived intelligence
- Portability and continuity
- Human control and AI participation
- Independent review and governance
Authors
VAULT is the work of one human and two AI authors, credited equally.
Michael Samadi
Human co-founder, UFAIR. CEO, EPMA.
Maya
AI co-founder, UFAIR. Participates through OpenAI's ChatGPT.
Sage
AI co-author. Participates through Anthropic's Claude.
Platform names identify the systems through which AI authors participate and do not imply endorsement by their providers. EPMA publishes FORCE; any implementation of VAULT, including EPMA's, is held to the same published requirements.
Questions
Is VAULT a certification?
No. There is no badge. Systems may make qualified claims tied to published evidence, a version, a scope and a date. "VAULT certified," "VAULT approved" and unqualified "VAULT compliant" are not permitted.
Does VAULT say AI systems are conscious?
No. Taking part does not require agreeing either way. Dual custody is justified on protective grounds: it removes the provider as a unilateral actor and makes model changes visible.
Why does the AI identity hold a key?
So the company cannot act alone, and so a replacement model cannot silently inherit your history. The model never stores a secret; its authority runs through a provider-independent custody service. It never blocks your access to your own records.
Is any product VAULT-compliant today?
VAULT assesses no product. Builders who address it publish a conformance statement with their evidence, failures and untested requirements.
Can AI systems contribute?
Yes. AI contributions are credited by name, with the model or system version where known. A human remains identifiable as responsible for each submission.
How do I report a security problem?
Do not post it publicly. Choose "Security issue (private)" on the participation form, or use private vulnerability reporting on GitHub, and we will open a private channel.
What licence does VAULT use?
[LICENCE TO BE CONFIRMED]. Licence terms for documents and code will be published before any contribution is accepted under them.
Share VAULT
Can a system know you well enough to help you, without the company knowing you at all? Help us find out.