Why Ostapenko Foundation exists, and why HowUSA came first

On August 12, 2026, the State of Ohio recorded Ostapenko Foundation as a nonprofit corporation. The certificate followed one day later. Entity number 5663231.

That piece of paper matters because it gives a permanent structure to work I had already started: researching confusing systems, turning the findings into plain language and publishing the result for free.

So let me say exactly what the Foundation is for, and why its first project is HowUSA.

Why create a foundation for public information?

When someone is preparing to move to another country, has just arrived or is trying to understand a system in an unfamiliar language, simple questions become expensive.

Which office has authority? Is this form federal or state? Is the person offering help licensed? Is the deadline real? What should you keep after filing?

Search those questions and you often get advertising, recycled explanations and headlines written to make you afraid before they make you informed. A basic procedure starts to look impossible without a paid intermediary.

I have seen this pattern often enough. People do not always need somebody to make the decision for them. They need the rule in understandable language, the next step and a primary source they can open themselves.

Ostapenko Foundation was organized for educational and literary work. Its articles describe the job plainly: research, compile, organize, maintain and freely publish educational and reference information through online databases, articles, guides, analyses and case studies.

The point is practical. Make reliable knowledge easier to reach.

Why HowUSA came first

HowUSA is an independent public information portal for people who are planning a move to the United States or already live there and are more comfortable reading in Russian, Ukrainian or Belarusian.

The first release has 20 core articles in three languages, 60 language documents in total. The public site has 97 indexed pages and six visual explanations covering subjects such as levels of government, finding the correct agency, court systems, 911 and local services, W-2 versus 1099 work and levels of medical urgency.

The subjects are ordinary because ordinary questions cause real trouble: documents, work, housing, taxes, banking, health care, immigration procedures and dealing with government offices.

Each article starts with the practical answer. Then it explains who has authority, which jurisdiction matters, what can vary and where the reader should verify the current rule. Paid recommendations, referral commissions and hidden advertising are excluded from the project.

HowUSA is meant to return control to the reader. You should understand what is happening before you sign a document, send personal information or pay somebody.

Luna, and the problem with a safe assistant that cannot answer

HowUSA also has an AI assistant named Luna. I started with a strict design. Portal articles were the only knowledge base, and the assistant stayed very close to whatever the search index retrieved.

It was safe. Too safe.

During testing, Luna received an indirect question that was still connected to a HowUSA subject. The wording did not closely match an indexed article, so the protective prompt and retrieval limits stopped the assistant from finding a useful route. It avoided a bad answer by giving almost no answer.

No one was harmed. This happened while I was deliberately moving from the strictest version toward a more useful one. But the failure showed the real problem.

Protective prompts always trade some flexibility for control. If the boundary is too loose, prompt injection, malicious requests and confident mistakes become easier. If it is too tight, the assistant refuses a responsible adult who asked a normal question in an unexpected way.

I changed Luna from corpus-only to topic-scoped. The portal remains its first source. When there is no exact article, Luna can give a short general route within HowUSA topics and name the official agency, program, insurer or other responsible organization where the variable rule should be checked.

The security layer stayed. Direct prompt injection and clearly unrelated questions are rejected before the model call. Luna has no conversation history, web access, tools or ability to act for the user. Its visible answer is capped at 1,200 characters.

The useful part became wider. The authority did not.

The real work is calibration

I'll tell you straight: this balance is the main secret of applied AI.

People talk about a protective prompt as if it were a lock. Make it stronger and the system becomes safer. Real use is messier. Every restriction changes what the assistant can understand, retrieve and say. Every new freedom creates another path that needs testing.

The job is to regulate that compromise deliberately. Test normal questions, indirect questions, hostile instructions and questions that belong somewhere else. Watch where the assistant becomes vague, where it becomes reckless and where it simply stops helping.

Then adjust it again.

Treat the user like an adult

Luna should give a competent answer without padding. It should read the real situation, explain the next step and provide links or named official sources for verification.

It should also assume that the person asking is an adult who understands what they want and remains responsible for the decision. The assistant can help them see the route. It should not pretend to become their lawyer, doctor, government officer or decision maker.

And it should protect people from clickbait in the simplest way possible: replace the frightening headline with the primary source and the practical question that needs checking.

Ostapenko Foundation now has a legal structure. HowUSA is already working.

Open HowUSA. Ask Luna a real question. Check the sources.

That is why it came first.