The organization
Alignment is not a specification problem.
The prevailing approach treats alignment as a matter of specification. Write down the right values, train them in, verify the result. Our research argues that this picture fails at its foundation, and that the failure is structural rather than a matter of engineering maturity. A specification fails at closure, the moment it stops updating from the process it governs. Optimization toward frozen objectives provably discards the path-dependent structure that values are made of.
The alternative is developmental. Values are processes, formed and maintained through feedback, resistance, and consequence. Velanox builds the formal groundwork, the empirical case, and the architectures in which values can develop rather than be imposed.
Velanox was founded in 2026 by Austin Spizzirri, who leads the research program.
The program
From diagnosis to foundations to construction.
Diagnosis
Formal results locating where specification-based alignment fails: the closure of static value specifications, the geometry of value collapse under substrate-independent optimization, and impossibility results for frozen, proxy-monitored systems.
Foundations
The empirical science of values. Developmental psychology, neuroscience, and comparative cognition converge on values as friction-dependent processes, with concrete architectural requirements for moral agency.
Construction
Subject Zero, a reference architecture for value emergence through embodied moral development, evaluated in pre-registered multi-agent testbeds.
Selected work
Published
The Specification Trap: Why Static Value Alignment Alone Is Insufficient for Robust Alignment
Five further papers in the series are in preparation.