Project summary

I'm building SEVERANT, an AI architecture designed from first principles rather than adapted from existing model paradigms. It's a full stack of seven core layers plus a bridge layer. A dedicated perception and actuation layer handles all input and output with the outside world, whether text, speech, or other modalities. A causal world-model runs an internal simulation environment to test hypotheses before committing to a decision. A tiered memory architecture represents long-term memory in natural language rather than opaque embeddings. A reasoning layer builds meaning through grounded conceptual composition rather than next-token prediction, working in continuous dialogue with the world-model. A bridge layer reconciles the world-model and reasoning layer's output into a single validated result before it's passed back to the perception and actuation layer.

The architectural property I'd point a reviewer to first is this: the safety evaluation layer's verdict is constructed independently by two separate reasoning layers, so that compromising one of them alone cannot produce a false verdict. That's a specific, testable structural claim, not a general assurance. The safety layer's predicates are formally specified and machine-checked using SVL, a verification language and compiler I wrote specifically for this purpose. A self-improvement layer continuously works to improve the system's own implementation, with no technical ability to modify the safety layer, enforced by the absence of a write path rather than a policy.

The architecture is fully specified. SVL's compiler is real and working, over 80 passing tests, LLVM IR generation, a GPU-targeting backend. A 41-scenario test suite, which I authored myself alongside the predicates it checks, currently passes in full, confirming the predicates behave as intended against the morally and ethically complex situations I constructed to test them. This is internal validation of design consistency, not independent adversarial red-teaming, and I want to be precise about that distinction rather than let the word "adversarial" imply more than it currently means. Independent adversarial evaluation, by people trying to break the independence property rather than confirm it, is exactly the kind of test this funding is meant to eventually make possible, once the system is trained and running rather than only specified.

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

The central hypothesis is that an AI system built with an explicit world-model, tiered natural-language memory, and a structurally independent safety layer outperforms current architectures on reasoning quality, safety, and interpretability, through structure rather than scale. Concretely, once trained, this is testable: whether internal simulation improves performance on reasoning tasks involving genuinely novel situations outside the training distribution, compared against baseline transformer models on the same tasks; whether natural-language tiered memory holds up on long-horizon retrieval consistency compared to embedding-based memory, measured directly rather than assumed; and most importantly, whether the safety independence property survives deliberate, external adversarial attack once qualified outside reviewers can actually attempt to break it, not just the self-authored scenarios described above. If the safety independence property fails under real external adversarial pressure, that's a negative result I'd report as such, not something I'd expect to quietly succeed by default. I'm building the architecture layer by layer against expanding real test coverage, rather than presenting a finished system and asking for funding to deploy it.

How will this funding be used?

This funding covers two things, in sequence, and is structured so a partial grant still buys something concrete rather than an unusable fragment. The first priority, and where funding below the full amount would go first, is relocation and living expenses, rent, moving costs, essential furniture, a capable laptop sufficient to move from planning into real implementation work immediately, and food, utilities, and other basic living costs, covering 12 months. This is included directly because most funding decisions take months to arrive, and full-time work on SEVERANT isn't possible without it in the meantime, regardless of what compute is eventually available. Once that's covered, the remainder is dedicated to a full workstation-class compute environment, capable CPU/GPU hardware, storage, and networking, sufficient to move from a laptop-scale bridge into real training loops and full-layer integration testing. This is the current hard technical bottleneck: the existing working environment cannot run real training loops or full-layer integration testing regardless of how complete the specification already is. Larger-scale dedicated hardware beyond this, GPU and FPGA servers for later-stage training and hardware validation, is being pursued separately and is not part of this request.

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

I'm a solo, self-taught researcher based in Cape Town, South Africa, with no formal degree or institutional affiliation, working on SEVERANT since November 2025. The technical foundation was built independently through university course material, primary technical documentation, and direct implementation work. What exists so far: a complete architectural specification across all seven core layers plus the bridge layer; a working Rust compiler for SVL with over 80 passing tests including LLVM IR generation and a GPU-targeting backend; real hardware design work including circuit-level compute-in-memory synthesis against an industry-standard process design kit and memory controller synthesis through an industry-standard toolchain; and the self-authored 41-scenario test suite described above.

I want to be direct about this rather than let it sit at the end of a paragraph: no formal external technical review of any of this has taken place yet. That's a real gap, not a caveat, and it's the main reason I'd welcome a reviewer here connecting me with someone qualified to look at the compiler, the hardware synthesis work, or the safety architecture specifically, more than I'd welcome the money alone. Everything, the code, the specification, and the synthesis results, is open to that kind of review rather than requiring trust in advance.

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

The most likely failure mode is stalling indefinitely at the design stage for lack of compute, which this request directly solves. Beyond that, the real risks are architectural, and I'd rather state them plainly than have a reviewer find them first: the safety independence property might not survive real external adversarial testing at scale, which would mean redesigning that layer rather than tuning it; several formal proof obligations related to the safety layer's guarantees remain genuinely open and might not resolve favorably; and the deeper hypothesis, that dedicated world-modeling and tiered memory meaningfully outperform current approaches, might simply turn out to be wrong once tested against real results. That's the honest range of outcomes for a real research question, not a settled conclusion dressed up as one. The project's longer-term hardware vision includes genuinely ambitious, industry-unproven components; none of that is depended on by the work this specific funding covers, which targets established, commercially available hardware.

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

None from external funders to date. All work so far has been self-funded on personal equipment. I applied to Foresight Institute's AI for Science & Safety Nodes program previously and was not selected. Since that application, the compiler has gone from partial to fully working with real passing tests, the hardware synthesis work has been completed, and the safety architecture's independence property has gone from a design intention to something validated, internally, against real constructed scenarios, which is specifically why I think this is worth a different outcome on reapplication rather than an unchanged one.