Project summary

IRL-2 reformulates the iterative reasoning core of a sparse Mixture-of-Experts architecture as a provably convergent fixed-point system, replacing a fixed number of refinement steps with a mathematical convergence guarantee. It's the direct successor to MoC (Mixture-of-Collaboration), a published architecture where routed experts collaborate through a learned mediator instead of merging outputs independently. In a matched 498.8M-parameter comparison trained on 9.83B tokens, MoC beat an identical MoE baseline on validation perplexity, 20.70 to 20.40, while improving expert load balance, with zero dead experts. IRL-2 is fully designed and implemented but has zero experimental validation. The one thing blocking it is compute.

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

Run the first controlled experiments validating IRL-2's core claims: that the reformulated reasoning core is a provable contraction, that it converges to a stable representation independent of iteration count, and that this improves on MoC's already-demonstrated gains. Method: implement the contractive iteration with spectral-norm-controlled layers, run matched ablations against the MoC v1 baseline at increasing scale, and measure both convergence behavior, contraction ratio and path-independence, alongside standard metrics like validation perplexity and routing health, using the same controlled methodology already used to validate MoC.

How will this funding be used?

Primarily cloud GPU compute: matched ablations from 64M to 500M parameters, replicating MoC's controlled methodology with added convergence diagnostics, on the same class of hardware, AMD MI300X, used for all prior results. Target around $30,000, roughly $18,000 in compute and $12,000 as short-term personal runway, so time currently spent on bug-bounty work, which self-funds this research today, can shift toward this instead.

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

Francisco Antonio, 18, self-taught, sole technical builder. I designed and trained MoC myself, including the training infrastructure, routing and capacity logic, and evaluation pipeline. Repo and full experiment logs are public at github.com/Auren-Research/lunaris. Before this, I built Lunaris Guard, a multilingual security classifier now in its third public version, and have an active bug-bounty track record across HackerOne, Immunefi, and Cantina. My co-founder Davi handles communication and operations. The technical work described here, including everything above, is mine.

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

Most likely failure mode is theoretical: with SwiGLU and RMSNorm inside the loop, a clean global Lipschitz bound isn't guaranteed the way it would be with simpler activations, so the contraction property could hold empirically at small scale but degrade at larger scale. A second, independent risk: even a genuinely convergent iteration might not outperform MoC v1 in practice, since recent work on weight-shared recurrent depth suggests convergence alone doesn't guarantee better task performance without added state capacity. If either happens, the honest outcome is a negative or mixed result, reported alongside the positive MoC v1 evidence rather than replacing it. Auren's broader work, Argus, MoC v1, and the security-model product, doesn't depend on IRL-2 succeeding to remain viable.

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

Zero outside funding in the last 12 months. Everything, compute, tooling, time, has been self-funded through paid security research, bug-bounty work across HackerOne, Immunefi, and Cantina.