Project summary

We started with a pretty simple question: why do some of the same problems keep showing up in AI systems in different ways?

Instead of treating each failure as its own separate problem, we started looking at what was underneath them and what they had in common. That led us to a growing set of invariants that seem to show up across different failure modes.

We've already tested this approach on specific problems. One example is the Samata Latch https://zenodo.org/records/21464368, a designed internal-state intervention for in-context binding integrity. The work also includes the experimental results, code, data, and replication materials https://zenodo.org/records/21479444. That work resulted in a paper accepted to the NeurIPS 2026 Workshop on Foundation Models and Language Model Security (FLMSec).

Now we're asking a bigger question. What if some of these problems aren't just things that need another safety mechanism added to the system? What if some of them are related to how the system is organized in the first place?

We don't know the answer yet. That's what this project is for. We want to take what we've learned so far, see what it actually implies at the architectural level, and then test it.

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

The first goal is to figure out whether the invariants we've identified actually tell us anything useful about how an AI system should be built.

We'll take the research we've already done and work through what those findings would require from an AI system. From there we'll develop architectural ideas that can actually be tested.

We're not starting with a finished architecture and trying to make the research fit it. The research comes first.

We'll build enough of the system to test the important ideas, run experiments, and compare the results with existing approaches. Some of the ideas will probably work better than others. Some may not work at all.

The point of this phase is to find out what holds up when we actually test it.

How will this funding be used?

The funding will mainly support the people doing the research and the technical work needed to test it.

That includes research and development time, experimental implementation, model and API access, compute, evaluation, analysis, and documenting the results.

At the $30,000 minimum, we'll focus on the core research question and the most important experiments. At the $50,000 level, we'll be able to test more variations, run a broader set of experiments, and spend more time on evaluation and replication.

This is a research phase. We're not asking for funding to build a commercial product or a large engineering team.

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

The project is being led by Creighton Baxter, working with a small AI research team based in Costa Rica and the United States.

The work comes out of several years of studying AI behavior, failure modes, security, and evaluation. The basic approach has been to look at a problem that keeps showing up, figure out what is actually underneath it, and then see if that underlying property can be tested or changed.

The best example so far is the Samata Latch. We started with a specific problem involving in-context binding integrity and developed a designed internal-state intervention to address it. We then tested it experimentally. The resulting paper, “The Samata Latch: Designed Internal State for In-Context Binding Integrity,” was accepted to the NeurIPS 2026 Workshop on Foundation Models and Language Model Security (FLMSec). The work also includes the experimental results, code, data, and replication materials.

That matters for this project because we're not starting from a purely theoretical idea. We've already gone through the process of identifying an AI problem, looking underneath the behavior, developing something testable, and seeing what the experiments tell us.

Now we want to apply that same process to the larger question of whether these underlying properties should affect how an AI system is organized.

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

The simplest possibility is that the invariants we've identified don't actually lead to a better architectural approach.

We could also find that an idea looks good in theory but doesn't make a meaningful difference when we test it. Or we could find that some of the assumptions behind the research were wrong.

If that happens, we'll document what failed and why. That's part of the research.

It could tell us that certain invariants aren't useful at the architectural level, that particular approaches don't work, or that we need to rethink some of the assumptions before going further.

We're not starting this project with the requirement that we produce a new architecture. We're trying to find out whether the research actually justifies one.

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

We have not raised outside funding specifically for this research during the last 12 months.

The work completed so far has been independently funded. That includes the research that led to the Samata Latch paper and the broader work on AI failure invariants.

We're now looking for outside funding to take that work into a more substantial experimental phase.

Please note

Automated tools can make judgments about how text was produced, but that isn't what this proposal is asking you to evaluate. The important questions are whether the research is real, whether the evidence holds up, and whether the results are useful. That's what we're asking people to fund.