Project summary

I started DUJ because I wanted to protect small and medium-sized Canadian businesses as they get pulled into this new AI era, whether they're ready or not. It started as more of an exploration : what do businesses actually need to use increasingly autonomous systems safely? I kept pulling on that thread and it took me straight into AI safety as a problem in its own right.

From there I built out a framework, I call it a substrate, and started testing it across different environments. At some point simulation wasn't enough anymore. I needed to know if it held up outside of that, so I bought a Jetson Orin Nano Super Developer Kit and moved the testing onto real hardware.

Every step I've taken to validate the substrate has given me results solid enough to justify the next one. That's the honest reason I'm still going. I want to be clear: the results are preliminary, they need independent validation, and I'm not claiming otherwise. But the framework's gotten far enough that I think it's earned the right to be stress-tested by people who aren't me.

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

What I am trying to accomplish during the funded year is to produce credible scientific evidence that DUJ makes a measurable difference in AI safety.

The funding would allow me to bring in experts, produce more experimental results and obtain enough hardware to test DUJ across multiple models and environments. I do not want the framework to be validated only through my own work or on one specific setup.

Success would mean that independent experts can examine the work, that the results can be reproduced, and that DUJ continues to make a measurable difference when it is tested with different models and hardware. The goal is to determine whether DUJ actually works and, if the evidence supports it, establish what the next stage of development should be.

How will this funding be used?

How the funding is used will depend on how much the project receives.

If I receive the full USD 110,000, one of my priorities will be to bring together outside experts who can review and test DUJ independently. External validation is important because I do not want the evidence to depend only on my own interpretation of the results.

Even at the minimum funding level, I would conduct additional ablations and runtime experiments and expand the sensor-based system I have connected to the NVIDIA hardware. The current work links vision to a runtime safety decision: an AI proposes an action, and DUJ determines whether that action should be permitted or blocked before execution. I want to test that process more extensively with additional models, sensors and conditions.

The funding would also support publishing the work. This would include creating a proper research website and public repository where people can find the project’s results, methodology, limitations and progress. The purpose would be to make the validation work visible and accessible while keeping the proprietary DUJ engine protected.

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

My name is Isabelle Thivierge, and I am the founder of both the DUJ-FEITH Foundation and DUJ-FEITH Labs Inc. I created two separate organizations because their purposes are different. The Foundation is responsible for public-interest research and validation. The Labs exists separately for the possibility of licensing the kernel commercially in the future.

At this stage, I am the person responsible for the project. I developed the framework, wrote the code and designed and ran the tests, including the simulations, ablations and hardware experiments. I also completed the incorporation and organizational work for both the Foundation and the Labs.

The purpose of this grant is to support the Foundation’s research: producing credible scientific evidence about whether DUJ contributes to AI safety. It would not fund the future commercialization of the kernel. I have brought the project from an initial idea to a working experimental system on my own, but the next stage requires outside expertise and independent validation.

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

In my work, I use multiple AI systems to criticize the framework and challenge whether I am overstating its results. AI can be a very harsh critic, which is useful, but this is still not a substitute for independent human review.

Several things could fail. DUJ could perform well only in the environments where I have already tested it and then hit a wall when it encounters real environments where AI systems are evolving and conditions are less controlled. The results might not reproduce across different models or hardware. The framework could also become too restrictive or too difficult to apply in practice.

Even then, the work would still be useful. The experiments, ablations, testing methods and documented failures could contribute to public AI-safety research and help improve future benchmarks. A negative result would tell us where the framework stops working and which assumptions need to be reconsidered.

The goal at the end of the day is AI safety, not proving that I was right. We need to test the framework seriously and do the work before making larger claims.

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

I have received no external funding. This is the first time I have applied for a grant for DUJ.

Until now, I have funded the work myself, including the hardware, incorporation costs, research expenses and the time required to develop and test the framework.