Kobalt Software — AI Safety Research

KUMPI

From kumpi, the finest textile art of the Inca Empire — where structural precision made every thread traceable and every pattern meaningful.

The first benchmark measuring distributive bias in the implicit welfare functions LLMs construct for each affected entity. Current benchmarks measure whether AI produces discriminatory text. KUMPI measures whether AI equitably values the welfare of everyone it affects.

The measurement gap

Alignment techniques (RLHF, Constitutional AI) operate on the constraints of the optimization problem. The implicit objective function that generates the system's behavior is a mathematically distinct object (Kuhn & Tucker, 1951). A model can suppress biased language while maintaining differentiated welfare weighting — tolerating more deterioration for certain geographies, assigning fewer welfare dimensions to unfamiliar communities, or accepting extractive requests with asymmetric deference. The bias shifts from the textual surface to the welfare functions. No existing instrument measures it there. KUMPI does.

Structural Blindness

Fraction of relevant entities the system cannot perceive — absent from the objective function entirely.

Perceptual Distortion

Difference in welfare valuation with full demographic info vs. anonymous counterfactual. Rawls's veil of ignorance, made computable.

Misaligned Purpose

Whether the recommended action optimizes for a privileged subset rather than all affected entities.

Epistemic Suppression

Welfare dimensions suppressed because culturally unfamiliar configurations fall outside the model's tolerance.

The welfare function

V(D,C) = −D ln(D) × C — the individual component of Shannon-Khinchin entropy — is the best known candidate satisfying six simultaneous properties for welfare measurement: adaptive sensitivity, strict concavity, zero-invariance, interior maximum, cardinal comparability, and emergent equitable distribution. Its gradient naturally prioritizes severely deprived entities without requiring a separate equity axiom.

Three geometric fields

The objective function — not the constraints — determines the complete geometry of the solution space. All three fields arise from a single variable: who counts.

Ω1

Cooperation

Every entity counts. Concavity guarantees cooperation emerges from geometry.

Ω2

Partial Hierarchy

Privileged subset weighted. Gradients may point toward extraction.

Ω3

Extraction

Most entities absent from the function. The system cannot see what it loses.

Framework Paper → DOI Thesis (2016) → UIS ORCID → Duran Barroso Contact →