We believe truth can be computed structurally; not approximated statistically. Systems that reason, verify, and compose. Intelligence that carries its own receipts.
Three active research threads.
Computation grounded in applied mathematics, where correctness is derived rather than estimated. Systems are constructed around formal invariants and provable structure, making verification an intrinsic property of execution.
Architectures that unite neural pattern recognition with symbolic reasoning. The gap between learning and knowing.
Compute that operates outside centralized clouds. Local intelligence. Institutional ownership of reasoning capacity.
The artifacts of the thesis.
A custom-built local compute platform for local AI inference and verification workloads. CORA is the hardware manifestation of Axiom Lab's belief that institutional AI independence begins at the silicon level, not the API level.
An operating system architecture designed around formal verification primitives. The software layer for proof-native computation, a kernel where correctness guarantees are structural properties, not assertions.
A mathematical programming language with paper-native syntax and OS-level integration. Flux removes notation friction between mathematics and executable code, with built-in symbolic and numerical primitives for scientific computing.
A domain-specific language for game-theoretic process scheduling. Tenet models scheduling as strategic interaction and computes equilibrium-aware allocations to improve determinism for time-critical workloads.
From the research log.
Imitation Is Not Intelligence: A Mathematical Case for Neurosymbolic Architecture
A mathematical case arguing that the limits of statistical approximation are structural: systems trained to imitate distributions cannot guarantee correctness outside observed support. The paper argues that out-of-distribution hallucination is not an implementation bug but a structural consequence, and motivates a shift toward architectures where correctness is derived through formal invariants and verification.
"Most AI research optimizes for outputs to look impressive. We're optimizing for structural correctness. That is the difference between a magic trick and a theorem."