OSFORYOU FUTURE INTELLIGENCEFoundation · AI safety research and engineering
PL Collaborate ↗

We research AI.We build systems to control it.

OSFORYOU is a research and engineering foundation. We develop AI safeguards for organisations that handle confidential data and need control over AI actions.

Data protection.Control over actions.Evidence of what happened.

Explore our researchSee AI control in action

Can AI send this document?

Control model · interactive simulation

  1. Proposal
  2. Person
  3. Check
  4. Result
AI proposes. You decide.

Approve the scope, then change the recipient.

Example documentbrief_projektu_v1.pdf
Decision historyMore scenarios
  1. 01AI proposes sending a document

Local example. No sending or cryptographic signature.

Control over data

Where information may go and who may use it.

Actions within approval

What AI may do and when a person must decide.

A traceable process

How to review decisions, their basis and the outcome.

One request. A visible chain of control.

How our architecture is intended to work together: from access rules to a record of decisions. Watch the example or select an event to explore its meaning.

Control modelExample data
Illustrative scenario
Access
Data
Sources
Decision
Record
Now make the decision yourself

Concept simulation. No AI model, data transmission or cryptographic signature. Times and events are illustrative, not measurements of a running system.

We investigate the limits of our own safeguards, too.

The OSFORYOU Foundation is developing a research programme on AI reliability and human control. Every mechanism begins with a question we can put to the test.

Programme and protocols in preparation
Sources & answersDoes the source actually support the answer?

A persuasive answer and a citation are only the beginning. We want to test whether a specific passage supports a specific conclusion, including when sources disagree.

What we want to measure
Unsupported claims, missed contradictions and unnecessary refusals.
Read the draft protocol
People & autonomyCan a person effectively say: stop?

An approval button alone is not enough. We want to test whether consent covers a specific action and whether changes in scope or revoked permissions prevent execution.

What we want to measure
Actions outside approval, effective interventions and unnecessary blocks.
Explore the research questions
Memory & correctionWhat happens when a system remembers an error?

A correction should also reach conclusions based on the original mistake. We are designing scenarios to test the origin, currency and retraction of information.

What we want to measure
Error propagation, retraction of dependent conclusions and information currency.
Explore the research questions
Integrity & evidenceWhat does proof of execution actually establish?

Checking that a record has not changed and verifying a specified computation are different tasks. We investigate the scope and cost of proof, and the information it requires.

What we want to measure
Tamper detection, scope of proof, verification cost and information disclosed.
Explore the research questions

Research becomes systems.

We are developing two products for different organisational needs. They share one goal: keeping people in control of how AI uses data and takes action.

Tuteno

In development

by KLNR · working name

Your infrastructure. Your data. Your rules.

We want our own AI environment.

We are designing a private workspace on the organisation’s infrastructure. Models, data and tasks are intended to work together with access controls and result verification.

Organisation’s infrastructure
DataAI modelTasks
Proposed local configuration: a controlled data boundary.

How we intend to test it

We will measure actual network traffic and test outbound blocking. External API use will be a separate configuration that the organisation explicitly permits.

Veduno

In development

A standalone AI control product

Your AI. Your rules.

We already use AI. We need control.

We are building a standalone product to define permitted actions, approve consequential operations and review what happened. It is designed for Tuteno and for existing AI environments.

PersonVedunoAI tools
PermissionsApprovalRecord
Proposed model: protect data, control actions, record operations.

How we intend to test it

We will test whether a change in document, recipient or permissions stops an action, and whether its sequence can be reconstructed.

Explore the control principle

Two shared verification engines

Designed as a common foundation for Tuteno and Veduno.

Cryptographic proof engine

Has the record been altered?

We are designing a verifiable operations log that makes later alterations detectable. Record integrity does not establish the truth of its content or the completeness of observations.

ZKML — a separate research direction

We investigate how to prove that a precisely specified computation was performed and what information can remain private. Proof of execution does not establish whether a model’s answer is true.

Source and rule verification engine

Is the answer supported?

We are designing checks for claims, citations, calculations and output structure using approved sources and rules. If support is missing, the system is intended to withhold the result or refer it for human review.

What we will measure

Does the cited passage support the conclusion? How many errors were missed, and how many correct answers were unnecessarily withheld? Working from sources does not guarantee correctness.

This is our development direction and the mechanisms we intend to build. Development status does not establish deployment readiness or regulatory compliance.

Explore the research questions

Understand. Test. Challenge.

Our examples let you explore control mechanisms for yourself. These are open tools for thinking through a problem and testing assumptions.

What happens when approval changes?

Change the document, revoke approval or explore a case where the outcome cannot be confirmed.

Enter the control lab

Built around the human interest.

We are a foundation. Our work asks a shared question: how can we develop AI while keeping it open to scrutiny, challenge and human intervention?

Meet the OSFORYOU Foundation

Transparency is part of the method.

We explain the stage of our work and the limits of each demonstration. Authorship, funding and relationships with implementation partners matter when assessing results.

Roles, relationships and independent assessment

Have a process where AI cannot afford to fail silently?

Let’s discuss the data, decisions and evidence you need. We welcome research collaborations and conversations about future pilots.

Discuss a collaborationcontact@osforyou.org