Kobalt Red Research · Kobalt Software LLC
A program that formally studies the objective functions AI systems build for affected entities — and constructs them with those whom no system listens to.
The program
The research unfolds in three interdependent movements. Each one changes the geometry of what rationality, entities, and alignment mean — not as constraints, but as structure.
Standard rationality (Cournot 1838, Nash 1950) is self-defeating: it produces results Pareto-dominated by cooperative solutions. The $R_S$ correction does not add constraints — it changes the geometry.
The correct unit is the entity, not the agent. To exist in the system is to have $W_e$, not agency. The Observational Closure Theorem founds external audit as logical necessity.
Alignment is not specified by rules. It emerges from which entities are included in the objective function and the properties of their welfare functions.
Geometry
The topology of inclusion determines the topology of outcomes. Three geometries arise from the welfare function $V(D,C)$ — not as categories imposed from outside, but as structures that emerge from the mathematics.
Pareto solutions emerge naturally. When all welfare functions are included and $V$ is strictly concave, cooperation is not a moral choice — it is the geometry.
Structured power with accountability. Partial inclusion with a monotone $V$ produces stable hierarchies — auditable when the Observational Closure condition holds.
Structural invisibility. Entities outside the objective function cannot be optimized for. Their exclusion is not a policy choice — it is a mathematical consequence.
Applications
Two direct applications derive from the three geometries without additional axioms. Both are currently in draft, with one featuring a live interactive demo.
Jensen inequality applied to $V(D,C)$: every system using averages incurs a quantifiable cost concentrated on the most vulnerable. Interactive demo with real IHME and IPCC data.
The PCC-W protocol for building welfare functions with communities, ecosystems, and future generations. Participatory construction — not extraction.
Programs
Three ongoing programs apply the framework to specific domains — each one a convergence of formal theory, real-world data, and communities whose welfare functions are rarely measured.
Invisibilized entities in clinical AI: Afro-descendant and indigenous communities, microbiomes, animals in research, rare conditions, LATAM and Global South health systems.
Ecosystems as entities with tipping-point welfare functions, future generations under the temporal factor, indigenous knowledge as formal input.
Current benchmarks measure textual output. KUMPI measures the implicit welfare function the model constructs for each affected entity.
We have spent years building conceptual roots so that growth, when it arrives, is coherent with what it claims to be.— Kobalt Red Research · About the program
About the program
Kobalt Red Research is a small, independent program inside Kobalt Software LLC. Its work is rigorous, unhurried, and oriented toward those whom formal systems tend to render invisible.
About the program →