Project summary

POFINO is an independent research and development project exploring how AI systems that interact with the physical world can remain under explicit human control.

The research focuses on separating what an AI may observe or propose from what it may store as approved knowledge or execute as a physical action. We want to investigate an authorization boundary supported by persistent memory and an inspectable audit trail.

We are seeking $10,000–$20,000 for a focused development and evaluation cycle. We are not presenting this as a validated safety system or a deployment-ready product.

What are this project's goals? How will you achieve them?

We propose to investigate three questions:

1. Can an AI proposal remain clearly separated from authorization to act?

2. Can persistent memory preserve the distinction between observations, model interpretations and human-approved decisions?

3. Can an audit trail make unauthorized actions and incorrect memory updates easier to detect?

The proposed approach is to define the permission and memory model, implement a limited demonstrator, and evaluate it first in software and then, where feasible, through controlled physical tests.

Test cases would include missing approval, ambiguous instructions, stale authorization, conflicting memory and model-generated claims presented as facts. We would document both successful controls and failures. Any physical testing would require a separately reviewed scope and explicit human authorization.

How will this funding be used?

Funding would support development time, model/API and compute costs, evaluation tooling, documentation, and components needed for a limited physical demonstrator.

At the $10,000 minimum, we would prioritize the authorization model, memory controls and software evaluation. Funding up to $20,000 would allow a broader validation cycle, including additional controlled physical testing and independent technical review where feasible.

These are proposed spending priorities. An itemized budget and deliverable scope would be agreed before accepting funding. The project has no fixed funding deadline.

Who is on your team? What's your track record on similar projects?

We are currently a two-person team: Nizam Yüksel and his sister.

I am Nizam Yüksel, an independent product developer in Türkiye. Since around 2015, I have worked across machine design, product development and manufacturing R&D. My recent work includes AI-agent workflows, organizational memory, explicit human approval and source provenance.

This experience provides practical context for investigating how AI-generated proposals can affect technical decisions and physical systems. It does not establish that the proposed safety architecture has already been validated. My sister's responsibilities and relevant background will be added before publication.

What are the most likely causes and outcomes if this project fails?

The main risks are that approval controls prove too cumbersome, permissions can be bypassed through an unexpected execution path, or persistent memory carries an incorrect assumption into a later decision.

A further risk is that successful software tests do not transfer to physical operation. We would therefore distinguish software results from physical test results and avoid broad safety claims.

If the approach fails, the useful outcome would be a documented account of failure modes, test cases and limitations. We would reduce or stop physical testing if the controls could not support the intended test scope.

How much money have you raised in the last 12 months, and from where?

We have received $0 in external funding for this project during the last 12 months.