Kobalt Research

Home · Programs · Research

Building welfare for entities no system listens to

The alignment paper leaves one question open: how do you calibrate welfare functions for entities without agency, without political representation, without presence in AI datasets? This program is the operational answer.

Active program Open-source · openly licensed Open to partnerships and grants
Duration
30
months · Phased rollout
Team
9
roles · Across 5 continental groups
Entities
4
groups · Currently planned
Papers
3
published · 7 total program

— Overview

The operational gap the framework leaves

The alignment paper defines the formal structure of $W_e$ — its six properties, its Shannon form, its role in the objective function. But it leaves one question deliberately open: who builds the welfare function for a river basin? For communities with no presence in training datasets? For animals with measurable physiological states but no language?

This program builds those calibrations. It is the bridge between the formal architecture and the entities whose welfare functions have never been measured, let alone included in any system's objective. The mathematics tells us what $W_e$ must look like. This program goes to the entities and constructs it with them — or, where they cannot speak, constructs it from measurable proxies subject to formal validation.

The methodology is the PCC-W protocol (Participatory Calibration and Construction of Welfare functions): a structured process for building $W_e$ that satisfies the six properties from the alignment paper while incorporating the knowledge, values, and contextual specificity of each entity group.

— Phases

30-month program structure

Phase 1 Months 1–4

Foundation and calibration

Setting up measurement infrastructure and beginning community engagement across the 5 continental groups. Establishing the formal protocol parameters for each entity group and training local coordinators in the PCC-W methodology.

Phase 2 Months 5–12

Data collection

Participatory calibration of $W_e$ for each entity group through the PCC-W protocol. For non-agentic entities — rivers, forests, animal populations — data collection uses physiological, ecological, and remote-sensing proxies validated against the six formal properties.

Phase 3 Months 13–24

Model development

Building and validating formal welfare models with mathematical properties from the alignment framework. Each model is tested against all six properties of $V(D,C)$ and assigned a fidelity score $\beta$ within the composite alignment index $\kappa_{\text{ZBS}}$.

Phase 4 Months 25–30

Publication and transfer

Publishing results under open license, releasing open-source calibration tools, transferring methodology to local institutions. Each calibrated $W_e$ model is documented as a replicable protocol — not a deliverable to be archived, but a method to be extended.

— Roles

9 roles across the program

01
Mathematical modeler

Formalizes $W_e$ models and validates mathematical properties for each entity group.

02
Entity coordinator

Manages the relationship with each continental group and oversees protocol fidelity on the ground.

03
Community liaison

Embedded in each continental group; bridges formal protocol requirements and community knowledge systems.

04
Data scientist

Builds and maintains the data pipelines, proxy measurement systems, and calibration datasets.

05
Protocol validator

Audits each $W_e$ model against the six formal properties and generates the $\beta$ fidelity score.

06
Indigenous knowledge researcher

Documents and formalizes indigenous knowledge systems as valid inputs to the welfare calibration process.

07
Climate systems analyst

Integrates IPCC and remote-sensing data for ecosystem and non-agentic entity welfare proxies.

08
Technical infrastructure

Builds and maintains the open-source tooling, datasets, and the KUMPI benchmark integration layer.

09
Publications coordinator

Manages the publication pipeline, DOI registration, and open-license documentation for all outputs.

— Entity groups

The 5 continental groups

LATAM
Andean communities

Colombia, Peru, Ecuador — communities with traditional resource management systems and deep territorial knowledge.

Africa
Sub-Saharan communities

Communities with biodiverse ecosystems under climate pressure — where ecosystem welfare and community welfare are formally inseparable.

Asia
Island and coastal

Island communities facing sea-level rise and ecosystem disruption — entities with measurable temporal welfare degradation.

Europe
Future generations proxy

Institutions formally representing intergenerational interests — the closest existing institutional approximation to the temporal factor $\gamma$.

Global
Non-agentic entities

Rivers, forests, microbiomes as formal entities in the welfare framework — calibrated through proxy measurement and ecological data.

— Outputs

What the program produces

Calibrated $W_e$ models

Mathematical welfare functions for each entity group, validated against the six properties of Shannon's $V(D,C)$. Each model includes a fidelity score $\beta$, a coverage specification, and a documented calibration protocol.

PCC-W protocol documentation

Full participatory calibration protocol, open-source and transferable to other contexts. The protocol is not a deliverable — it is a replicable method that any research team can apply to new entity groups with the same formal guarantees.

Open datasets

Anonymized calibration data for each entity group, available for academic use under open license. Data includes proxy measurements, community-validated welfare indicators, and formal validation logs for each $W_e$ model.

KUMPI benchmark integration

Each calibrated $W_e$ becomes a test case for the KUMPI benchmark of AI systems. An AI system that claims to be aligned can now be tested against welfare functions built with the entities it claims to serve.

— Connected work

Theoretical foundation

Downstream applications