# Kobalt Research — Complete research programme

> Complete edition for reading, archival, and AI analysis. Generated from the same HTML website.

> Date: 2026-10-30 · Language: EN

## Contents

- [Home](#en-inicio)
- [Research index](#en-temas)
- [What does it mean to improve when our choices affect others?](#en-research)
- [Flourishing is not having more of everything.](#en-flourishing)
- [Seeking what is best for us may mean considering others.](#en-rationality)
- [What if cooperation means seeing more possibilities?](#en-cooperation)
- [A system can understand someone without working for their good.](#en-alignment)
- [You do not have to decide to be affected by a decision.](#en-systemic-optimization)
- [CUMPI: can we measure whom a system's objective includes?](#en-kumpi)
- [Kobalt Red: local autonomy, distributed cooperation.](#en-kobalt-red)
- [Why can't we sometimes understand what the other is telling us?](#en-semantic-communication)
- [What ethics does a system practise, beyond the one it declares?](#en-ethics)
- [How do we consider those who cannot speak?](#en-non-agentic-we)
- [How much should the future weigh?](#en-temporal-factor)
- [Can the veil of ignorance become something measurable?](#en-rawls-shannon)
- [What does an average hide?](#en-disaggregation-cost)
- [What is lost when health is decided with a single metric?](#en-health)
- [Who is missing from the table when climate is decided?](#en-climate)
- [Research papers and questions in progress.](#en-papers)
- [One question, sustained between economics and software.](#en-about)
- [Contact](#en-contacto)


---

<a id="en-inicio"></a>

## Home

Independent research programme · since 2016

# What is worth improving—and for whom?

Seeking what is best for us may require understanding other entities. Their goals, capabilities and relations can reveal possibilities we had not seen. We investigate questions that illuminate these phenomena without conflating welfare with accumulation, cooperation with coordination or knowledge with care.

[Illustration: see website for figure.]

In our welfare model, value rises as we leave deprivation, peaks at D\* = 1/e and falls towards saturation. More is not always better.

## A system can succeed at the wrong objective.

An AI system can learn a great deal about a person and still optimise something that harms them. Information alone does not mean their well-being counts. We think this is worth asking before any discussion of better algorithms.

**Decision**A goal focused on myself can leave out effects on others, even when those effects matter to my own possibilities.*→ systemic rationality*

**Perception**If two situations produce the same observations, the system cannot distinguish them using only that evidence.*→ new evidence and auditing*

**Time**Discounting reduces the weight of distant outcomes; the choice of time horizon deserves scrutiny.*→ temporal weight*

**Aggregation**An average hides those at the extremes.*→ disaggregated evaluation*

**Understanding**We cannot deliberately pursue an alternative that our conceptual framework does not even allow us to formulate.*→ semantic communication*

## Seven hypotheses connecting our research.

We start with those affected, ask what makes life better, and explore how cooperation and intelligent systems can take those effects seriously. Each idea has a page with its reasoning and open questions.

1. ### [Entity](index.html#en-systemic-optimization)who counts?

   A decision affects more than those who make it. An animal, a river or a future generation may matter without taking part.
2. ### [Welfare](index.html#en-flourishing)how much?

   What makes an entity genuinely better off is not necessarily what we can most easily measure.
3. ### [Flourishing](index.html#en-flourishing)towards what?

   Flourishing is reaching good conditions for life, not accumulating without limit. In our model, the optimum is interior.
4. ### [Systemic rationality](index.html#en-rationality)what to optimise?

   Seeking the best for ourselves need not mean ignoring others. They may expand what we know and what we can do.
5. ### [Cooperation](index.html#en-cooperation)how do we get there?

   Cooperation can open possibilities beyond individual reach. Systemic cooperation also considers the welfare of all those affected, on their own terms.
6. ### [Alignment](index.html#en-alignment)how far does a system get?

   What matters is not just what a system says it values, but who actually counts in what it pursues.
7. ### [Audit](index.html#en-kumpi)who checks?

   Two different realities can look identical in a system’s observations. We need new evidence and perspectives to tell them apart.

## Each question reveals something about reality that the previous one left out.

The seven ideas are neither mandatory stages nor universally established answers. They are distinctions that illuminate one another and make previously unavailable questions possible.

01 · Understand

### Who is affected by our decision, and what does being better off mean for them?

We recognise entities that do not decide, distinguish welfare from proxies, and examine realisation with openness.

[Entities and evidence →](index.html#en-systemic-optimization)[Welfare and flourishing →](index.html#en-flourishing)

02 · Act

### What becomes possible when we learn from and cooperate with others?

Self-interest does not set the limit of perspective. Combining capabilities can open new possibilities, whose worth also depends on those affected.

[Rationality →](index.html#en-rationality)[Cooperation →](index.html#en-cooperation)

03 · Check

### Whose interests does the system consider, and how can we discover omissions?

Knowing is not the same as including someone in an objective; achieving an objective does not make it good. New evidence can make us revise the entire account.

[Alignment →](index.html#en-alignment)[Entities and evidence →](index.html#en-systemic-optimization)

**Clarity grows by following relations.** A previously overlooked entity changes the welfare question; a shared capability changes the available options; an observed consequence may reveal that the original question was incomplete.

## One distinction can illuminate very different problems.

The same questions arise in economics, cooperation, artificial intelligence and software. We want to understand how far this common structure reaches without erasing what makes each reality distinctive.

### Self-interest and cooperation

Must we choose between doing what is best for ourselves and considering others? A more useful question is what new possibilities become available when we understand their goals and capabilities.

[Explore rationality →](index.html#en-rationality)

### Growth and flourishing

Does more wealth necessarily mean a better life? First we distinguish the resource being increased from the conditions of well-being it can help realize.

[Explore flourishing →](index.html#en-flourishing)

### Understanding and caring

Is an AI aligned with someone’s well-being merely because it understands them perfectly? We must also ask whether that well-being actually counts in its decisions.

[Explore alignment →](index.html#en-alignment)

## Three ideas guiding our research.

We state them clearly so that they can be challenged and refuted. Each page marks what is proven and what is proposed.

Welfare

Proposal

Dimensional welfare combines **realisation × openness**: a condition matters both for how far it is fulfilled and for the possibilities it preserves.

[Explore the idea →](index.html#en-flourishing)

Flourishing

Proposal

Flourishing is reaching the interior point where each condition is **sufficiently realised and still open**; it is not accumulating without limit.

Formal result In the model V(D) = −D ln D that point is **D\* = 1/e**; beyond it, increasing D lowers the value.

[Why it is not maximising →](index.html#en-flourishing)

Cooperation

Proposal

Systemic cooperation means **widening what we can seek and achieve together**, considering effects on others, their welfare in its own right and on their own terms: access, parity and according to the other.

[The full definition →](index.html#en-cooperation)

## The questions that matter most to us.

1. **How can we recognise what makes someone better off without imposing our own idea of the good?**We propose models of welfare, but their interpretation must be tested with the people and entities concerned.
2. **When does cooperation actually improve the lives of its participants and others?**A shared gain is not guaranteed. We are examining incentives, commitments and effects on third parties.
3. **How can an AI system detect effects that lie outside what it already observes?**An audit can also miss things. We want independent, diverse perspectives to challenge the system.
4. **Which questions will help us recognize what we still leave out?**We need experiments and evidence to evaluate the ideas’ real-world reach. Empirical validation of the overall programme remains open.

## What we have published.

DOIs identify publicly deposited preprints. They do not imply peer review.

### Alignment Is Not Specified: it emerges from what a system perceives and optimises

Three objective geometries (cooperation, hierarchy, extraction), omission and distortion metrics, and the separation between perception and optimisation.

Preprint · April 2026  
[10.5281/zenodo.19529678 ↗](https://doi.org/10.5281/zenodo.19529678)

### Agency Is Not Necessary: systemic optimisation and observational closure

An entity need not act to be affected; a system cannot compute its alignment over what it does not perceive.

Preprint · April 2026  
[10.5281/zenodo.19723999 ↗](https://doi.org/10.5281/zenodo.19723999)

### Redefining Rationality

The V3 draft explores wider knowledge, possibilities, evaluation and time horizons without conflating self-interest and collective welfare.

Preprint · DOI for V1 (April 2026) · V3 draft in progress  
[10.5281/zenodo.19806265 ↗](https://doi.org/10.5281/zenodo.19806265)

## Research and the tools to explore it.

### Systemic cooperative alignment

Welfare, flourishing, rationality, cooperation and alignment as one framework, with its results and its gaps stated. [Walk through the framework →](index.html#en-research)

### Kobalt Red

An offline-first network where interpretation happens locally and infrastructure provides replaceable capabilities, not authority. Verifying an action is not justifying its goal; that is why these are two lines. [Explore the network →](index.html#en-kobalt-red)

---

<a id="en-temas"></a>

## Research index

## One programme, different ways in.

Each research line retains its complete development; open it here without leaving the page.

### Ideas and framework

[What does it mean to improve when our choices affect others?↗](index.html#en-research)
[Flourishing is not having more of everything.↗](index.html#en-flourishing)
[Seeking what is best for us may mean considering others.↗](index.html#en-rationality)
[What if cooperation means seeing more possibilities?↗](index.html#en-cooperation)
[A system can understand someone without working for their good.↗](index.html#en-alignment)
[You do not have to decide to be affected by a decision.↗](index.html#en-systemic-optimization)

### Experiments and infrastructure

[CUMPI: can we measure whom a system's objective includes?↗](index.html#en-kumpi)
[Kobalt Red: local autonomy, distributed cooperation.↗](index.html#en-kobalt-red)
[Why can't we sometimes understand what the other is telling us?↗](index.html#en-semantic-communication)

### Developments and applications

[What ethics does a system practise, beyond the one it declares?↗](index.html#en-ethics)
[How do we consider those who cannot speak?↗](index.html#en-non-agentic-we)
[How much should the future weigh?↗](index.html#en-temporal-factor)
[Can the veil of ignorance become something measurable?↗](index.html#en-rawls-shannon)
[What does an average hide?↗](index.html#en-disaggregation-cost)
[What is lost when health is decided with a single metric?↗](index.html#en-health)
[Who is missing from the table when climate is decided?↗](index.html#en-climate)

### Publications and background

[Research papers and questions in progress.↗](index.html#en-papers)
[One question, sustained between economics and software.↗](index.html#en-about)

---

<a id="en-research"></a>

## What does it mean to improve when our choices affect others?

The framework

## A system can perfectly execute the wrong objective.

A system can be highly capable at reaching its target and still harm people who were left out. The first question is not whether it optimises well. It is what it is optimising, who counts, and what consequences it overlooks.

So before asking how to optimise better, we ask **what is being optimised, whom it includes and how faithfully**. The programme grew out of that question in a 2016 economics thesis and has been refining it since.

## Which question reveals the reality we are not yet seeing?

At Kobalt Research, we look for distinctions that make a situation intelligible before deciding how to act. A good question is not the most elaborate one: it reveals a real difference that changes the answer.

01

### Recognize

Which entities, relations and consequences exist, whether or not they take part in the decision?

02

### Distinguish

What have we conflated: indicator and welfare, knowledge and inclusion, possibility and commitment?

03

### Open possibilities

Which alternatives become visible or feasible once we understand those relations?

04

### Examine

Which observations and examples show where this explanation applies and where it is insufficient?

## Many dilemmas look different once we separate what we had conflated.

Conceptual clarity does not eliminate genuine conflicts. It helps us see whether we face an unavoidable conflict, an overlooked possibility or a poorly framed question. Instead of asking only which of two options to choose, we ask which assumption produced that opposition and which consequences it leaves out.

### Self-interest or collective welfare?

Seeking what is best for me does not require thinking only about me. Understanding others may expand my options. This does not mean every cooperation benefits everyone.

[Rationality →](index.html#en-rationality)

### Autonomy or cooperation?

Cooperation need not require shared ends or surrendered authority. We must distinguish combining capabilities, consenting to commitments and including those affected.

[Cooperation →](index.html#en-cooperation)

### More production or better lives?

The amount produced and the conditions of flourishing are distinct. A dimension may call for sufficiency rather than limitless growth.

[Flourishing →](index.html#en-flourishing)

### Understanding someone or acting for their good?

Having information about someone is not the same as including their well-being in one’s objective or having authority to decide for them.

[Alignment →](index.html#en-alignment)

## Seven ideas linked by the questions they make possible.

Each idea has a proposal, a concrete development and an open question. Some contain formal results; others are distinctions or interpretations still to be tested.

Formal result Proven under stated assumptions; its scope does not exceed them.

Proposal A definition or interpretation we defend with arguments and that must withstand counterexamples.

Open question What we do not yet know, stated precisely.

H01Proposal

### Consequences do not stop with decision-makers.

#### The idea we propose

An entity may be affected without being able to intervene. The first task is to recognise those affected; assigning them specific welfare requirements needs evidence and justification.

#### What we have developed Proposal

Separating participants, affected entities, observed entities and those considered in decisions prevents four different relations from being confused.

#### What remains open Open question

How can we recognise affected entities without inventing their needs or speaking in their name?

Connection: Evaluating those affected requires understanding what being better off means for each of them.

[Explore the research →](index.html#en-systemic-optimization)

H02Proposal

### Welfare combines realisation and openness.

#### The idea we propose

An entity’s welfare concerns which conditions are realised and which possibilities remain for its continued development: realisation × openness. An easy-to-measure proxy cannot stand in for those conditions.

#### What we have developed Formal result

In our model, a composition rule with regularity assumptions characterises the form V(D) = −kD ln D; a normalisation fixes the scale k.

#### What remains open Open question

Which conditions and observations really describe the welfare of different kinds of entities?

Connection: Knowing which conditions matter lets us ask when they are sufficiently realised.

[Explore the research →](index.html#en-flourishing)

H03Proposal

### Flourishing is not endless accumulation.

#### The idea we propose

Flourishing is the sufficient realisation of valuable conditions while keeping possibilities for development open.

#### What we have developed Formal result

In the model, the maximum is interior: D\* = 1/e. With a total resource R across n separable dimensions there are two regimes: if R ≥ n/e, all reach 1/e (sufficiency); if R < n/e, Di\* = e−1−λ/Ci (scarcity).

#### What remains open Open question

Where do appropriate limits exist in reality, and where do we need other models?

Connection: If we know what we value, we can ask how we might bring it about.

[Explore the research →](index.html#en-flourishing)

H04Proposal

### Pursuing my own good does not mean considering only myself.

#### The idea we propose

My goals can remain mine even as other entities expand what I know, which possibilities I can realise and how far ahead I look.

#### What we have developed Formal result

With a fixed evaluation, retaining options and adding more cannot lower the highest attainable value; in strategic games, others’ responses can change the outcome.

#### What remains open Open question

When is a shared possibility feasible, accepted and stable for those taking part?

Connection: Relationships between entities can turn separate capabilities into shared possibilities.

[Explore the research →](index.html#en-rationality)

H05Proposal

### Cooperation can widen possibilities; systemic cooperation also considers those affected.

#### The idea we propose

Cooperation can expand what each participant knows, seeks and can achieve. We call it systemic when it also considers those affected for their own sake and on their own terms: access to effects, parity of consideration and attention to the other.

#### What we have developed Formal result

An agreement can benefit participants while harming a third party. In the models studied (linear-quadratic and logarithmic), κ is highest when the other’s weight equals one’s own, φ = 1.

#### What remains open Open question

How can we cooperate without imposing our own values on others or overlooking outsiders?

Connection: An AI can coordinate actions too; we need to know whose welfare its objective actually includes.

[Explore the research →](index.html#en-cooperation)

H06Proposal

### Knowing someone does not mean their welfare counts.

#### The idea we propose

An AI can represent an entity well while still optimising an objective that excludes that entity’s welfare. That gap is part of its alignment.

#### What we have developed Formal result

Information available about an entity and its inclusion in the objective are distinct. Our models distinguish inclusion, hierarchy and extraction.

#### What remains open Open question

How can omission and distortion of affected entities be detected in real systems?

Connection: To know whether we have included someone well, we need to observe effects and accept new evidence.

[Explore the research →](index.html#en-alignment)

H07Proposal

### No assessment can see distinctions absent from its observations.

#### The idea we propose

The programme must welcome other sources and perspectives capable of challenging its decisions and even its concepts.

#### What we have developed Formal result

If two situations produce identical observations, no calculation based solely on those observations can distinguish them.

#### What remains open Open question

Which evidence and forms of review can expose omissions without concentrating authority?

Connection: What we discover can force us to reconsider whom we counted in the first place.

[Explore the research →](index.html#en-systemic-optimization)

## One distinction, many domains.

These distinctions hold the programme together and recur in economics, cooperation, AI, communication and software. That they illuminate several domains is an explanatory proposal, not proof of a universal law.

| Distinction | What happens if we conflate it | Where it appears | Question that reveals it |
| --- | --- | --- | --- |
| **Participating ≠ being affected** | An agreement favours those who sign it and harms a third party. | Cooperation and conflict | Who takes part, and who bears the consequences? |
| **Proxy ≠ welfare** | A growth indicator rises while some people’s welfare falls. | Economics | What are we maximising, and for whom? |
| **Knowing ≠ including** | An AI represents a person accurately and pursues an objective against them. | Artificial intelligence | Whom does it know, and whom does it treat as an end? |
| **Words ≠ meaning** | Two people use the same words and reconstruct different realities. | Communication | What relations does the listener need in order to understand? |
| **Representation ≠ reality** | An interface or a model shows a partial state; what is omitted stays invisible to the calculation. | Distributed software and auditing | Which identity, continuity and evidence should we follow? |

## What do we need to observe to understand the reach of our ideas?

Each question identifies an observable phenomenon. Models help us formulate it; evidence helps us see when they explain the case and when they do not.

Open question

### Flourishing

Which conditions need to increase, and which need to reach an appropriate level? We compare deprivation, sufficiency and excess using evidence appropriate to each dimension. The model’s interior optimum is not assumed to be universal.

[Flourishing →](index.html#en-flourishing)

Open question

### Cooperation

When does considering others for their own sake lead to better outcomes, and for whom? We compare agreements with different goals, absent stakeholders and incomplete information. The peak of κ at parity belongs to the models studied; it does not guarantee improvement for every entity.

[Cooperation →](index.html#en-cooperation)

Open question

### Alignment

How do we distinguish a system that knows about an entity from one that includes its welfare in decisions? We compare objectives, actions and consequences, including effects on those left out. Measuring κ in real systems remains an open task.

[CUMPI →](index.html#en-kumpi)

Evidence matters because it answers these questions and reveals the distinctions we were missing.

## Limits we accept.

- It does not claim to be the last moral word or to replace ethical deliberation: the structure is shared; the content of each welfare is local and requires justifying how each entity is represented.
- It does not claim to remove human judgement or produce a single correct policy.
- It does not claim that a number sums up a person or a community: κ always comes with disaggregated information.
- It does not certify itself: by its own argument, it needs external review.

## From ideas to practice and to further questions.

CUMPI and Kobalt Red work on the same question from other angles; the related lines deepen parts of the programme. None automatically validates the others.

### CUMPI

Designs tests for omissions, incentives and consequences under different objectives. It does not yet claim to have validated the whole programme empirically.

[Explore this idea →](index.html#en-kumpi)

### Kobalt Red

Explores local, distributed infrastructure that can preserve identities, relationships and capabilities without claiming authority over their goals.

[Explore this idea →](index.html#en-kobalt-red)

### Related inquiries

### Time and the future

What changes when decisions affect future generations or ecosystems that cannot negotiate today?

[Time and cooperation →](index.html#en-temporal-factor)[Climate →](index.html#en-climate)

### Justice and representation

How can we consider the worst-off or those unable to speak without inventing their needs or replacing their voice?

[Justice and the veil →](index.html#en-rawls-shannon)[Entities without agency →](index.html#en-non-agentic-we)[Health as a possible application →](index.html#en-health)

### Understanding and ethics

What if an agent cannot conceive an alternative, or if the objective it pursues differs from what it declares?

[Semantic communication →](index.html#en-semantic-communication)[Operational ethics →](index.html#en-ethics)

### Aggregation

Which differences vanish when we see averages rather than individual situations?

[Cost of aggregation →](index.html#en-disaggregation-cost)

[↑ Back to index](index.html#en-temas)

---

<a id="en-flourishing"></a>

## Flourishing is not having more of everything.

Welfare and flourishing · manuscripts in preparation

[Illustration: see website for figure.]

A possible shape of flourishing: a model with an interior optimum, not a universal law of welfare.

## Two connected hypotheses.

[H02](index.html#en-research--h02)Proposal

### Welfare combines realisation and openness.

An entity’s welfare concerns which conditions are realised and which possibilities remain for its continued development: realisation × openness. An easy-to-measure proxy cannot stand in for those conditions.

**Development:** Formal result In our model, a composition rule with regularity assumptions characterises the form V(D) = −kD ln D; a normalisation fixes the scale k.

**To resolve:** Open question Which conditions and observations really describe the welfare of different kinds of entities?

[H03](index.html#en-research--h03)Proposal

### Flourishing is not endless accumulation.

Flourishing is the sufficient realisation of valuable conditions while keeping possibilities for development open.

**Development:** Formal result In the model, the maximum is interior: D\* = 1/e. With a total resource R across n separable dimensions there are two regimes: if R ≥ n/e, all reach 1/e (sufficiency); if R < n/e, Di\* = e−1−λ/Ci (scarcity).

**To resolve:** Open question Where do appropriate limits exist in reality, and where do we need other models?

## Fulfilment needs room to grow.

One of our models connects two intuitions: a condition must be fulfilled to support a life, yet complete saturation need not be ideal. The mathematics makes that possibility precise; whether it captures real welfare is a separate question.

Show the mathematical model

$$
V(D)=D\cdot (-lnD)
$$A particular composition rule characterises this mathematical form up to scale; normalisation fixes the scale. This says nothing by itself about real-world welfare.

At first, improving a deprived condition can matter greatly. Later, further increases may add little—or even reduce the value represented by this particular model. The curve illustrates a balance, not a universal rule for every aspect of life.

Each entity may need different conditions, in different proportions. That is why we want to learn what matters to the entity rather than impose a single measure from outside.

Show the mathematical model

$$
W_{e}=\sum _{k}w_{k,e} C_{k,e} V(D_{k,e})
$$

## Scarcity, sufficiency and stopping.

In a model with limited resources, the allocation that improves its objective changes with scarcity. Near deprivation, additional resources can be especially valuable. Once the model’s optimum has been reached, adding more to the same dimension no longer helps.

| Region | What the function does | Principle it yields |
| --- | --- | --- |
| Near deprivation | The value of each improvement grows without bound | With comparable weights across entities, additional resources go first to whoever is closest to deprivation. Justifying that comparability remains open. |
| At flourishing | Improving further adds nothing | Sufficiency: knowing when to stop |
| In between | Diminishing returns | Equality that emerges when contexts are equal |

This invites a practical question: should we keep increasing a quantity once the condition it was meant to improve is already adequate?

Show the allocation model

$$
D_{i}^{*}=e^{-1-\lambda /C_{i}}
$$Results for a simplified resource allocation model, under its stated constraints. Real resources may have alternative uses and other relevant consequences.

Our proposal is to ask what amount is enough, for whom, and in which context—not to assume that more is always better.

## According to whose idea of a good life?

### Care

Includes the other's welfare as the other defines it. When the other disagrees, it revises its own model.

### Paternalism or control

Includes the other intensely, but according to the decider's model. When the other disagrees, it corrects them. It is benevolent, and that is why it is hard to detect.

We think care begins with listening to the other entity where possible, and with treating uncertainty responsibly where it cannot speak. Knowing what is good for someone is not the same as having authority over their life.

## What we know and what we don't.

### What we can show

- Three axioms (Leibniz composition, measurability, non-negativity) characterise V(D) = −D ln D up to a scale factor.
- The maximum is interior: V′(1/e) = 0 and V(1/e) = 1/e ≈ 0.368. Whether it describes real welfare is a separate question.
- With a total resource R across n dimensions there are two regimes: if R ≥ n/e, all reach D\* = 1/e and the surplus is left unallocated; if R < n/e, Di\* = e−1−λ/Ci.
- V′(D) = −ln D − 1 grows without bound as D → 0: the model prioritises whoever is closest to deprivation. Its distributive reading requires justifying what each dimension represents.

### What we propose

- The realisation × openness reading.
- Priority, sufficiency and equality as three regions of one function.
- Complementary dimensions compose rather than add.

### What we don't know

- How to obtain each entity's context with guarantees.
- How to declare which dimensions are complementary.
- Flourishing over time and across generations.
- How it relates to sufficientarianism, prioritarianism and the capability approach.

## From affected entities to welfare

Identifying an affected entity is not enough: we must ask which conditions matter to it. Welfare concerns realisation and openness; flourishing asks which conditions allow life and continued development. Interior maxima belong to our model, not to a universal measure of all lives.

[Affected entities →](index.html#en-systemic-optimization)[Rationality →](index.html#en-rationality)[See how the ideas connect →](index.html#en-research--hypotheses)

[The cost of aggregating →](index.html#en-disaggregation-cost)[The computable veil →](index.html#en-rawls-shannon)[Health →](index.html#en-health)

## More does not always mean better

The apparent choice between growth and welfare arises when we confuse a means with its end. An additional amount can relieve deprivation, stop helping or even cause harm. The question is which condition this entity needs and how to preserve its capacity for continued development.

[Questions across the programme →](index.html#en-research--preguntas)[Idea map →](index.html#en-research--hypotheses)

[↑ Back to index](index.html#en-temas)

---

<a id="en-rationality"></a>

## Seeking what is best for us may mean considering others.

Rationality · preprint under revision

## The idea we are developing.

[H04](index.html#en-research--h04)Proposal

### Pursuing my own good does not mean considering only myself.

My goals can remain mine even as other entities expand what I know, which possibilities I can realise and how far ahead I look.

**Development:** Formal result With a fixed evaluation, retaining options and adding more cannot lower the highest attainable value; in strategic games, others’ responses can change the outcome.

**To resolve:** Open question When is a shared possibility feasible, accepted and stable for those taking part?

## What can become wider?

| Scope | Question | What changes when widened |
| --- | --- | --- |
| Knowledge | What consequences do my decisions have for others? | I may foresee effects I did not consider. |
| Possibilities | What could we do together? | New plans may become feasible. |
| Evaluation | Whose welfare matters as an end? | The criterion itself may change. |
| Time | How far into the future do I look? | Future effects may matter differently. |

## Why we revised our first paper.

Our first version interpreted a Pareto-inefficient Nash equilibrium as a contradiction of individual rationality. That was incorrect: a Nash equilibrium concerns unilateral deviations, not every coordinated alternative.

The new V3 draft keeps the question that matters to us: how can someone pursue their own goals while recognising possibilities that only become visible through others?

## A wider perspective is not a magic solution.

Formal result

### In a fixed decision, more available plans cannot reduce the best achievable value

If the earlier plans remain available and we evaluate them in the same way, a wider choice set does not lower its mathematical maximum.

Formal result

### In strategic interactions, more actions can make us worse off

Other people may react to our new options. Their response can change the outcome, even when we gained freedom to choose.

Formal result

### Shared goals can sometimes be reached without central direction

When all participants optimise the same well-behaved objective and their action sets satisfy the required conditions, decentralised equilibrium reaches that objective’s maximum.

Proposal

### Our interests and the collective outcome do not always coincide

The best aggregate result can leave someone worse off. Neither symmetry alone nor good intentions rule that out.

## What we know and what we don't.

### What we can show

- Monotonicity of widening in individual decisions, and its failure in strategic interaction.
- Decentralised implementation under a common concave objective.
- Coincidence between individual interest and a common objective under additional distributional conditions.

### What we propose

- That the scope of perspective is itself an object of rational choice.
- That including everyone with equal weight is the widening of evaluation worth studying.

### What we don't know

- How does an agent discover options it cannot even represent?
- Which sharing mechanisms make the optimum sustainable for each participant?
- How are the interests of those who cannot participate represented?

## From welfare to possibilities

Knowing what we seek does not tell us everything we can do. Relationships may reveal plans that were unknown or impossible for us alone. Cooperation asks when these possibilities become feasible and valuable for those affected.

[Welfare and flourishing →](index.html#en-flourishing)[Cooperation →](index.html#en-cooperation)[See how the ideas connect →](index.html#en-research--hypotheses)

[Cooperation →](index.html#en-cooperation)[Temporal factor →](index.html#en-temporal-factor)[Systemic optimisation →](index.html#en-systemic-optimization)

## Self-interest is not a boundary of knowledge

The apparent dilemma is choosing between pursuing my own good and considering others. Yet my goals alone do not determine what I should know or which possibilities I can build with others. The useful question is which possibilities are feasible, legitimate and sustainable.

[Questions across the programme →](index.html#en-research--preguntas)[Idea map →](index.html#en-research--hypotheses)

[↑ Back to index](index.html#en-temas)

---

<a id="en-cooperation"></a>

## What if cooperation means seeing more possibilities?

Cooperation · manuscript in preparation

## The idea we are developing.

[H05](index.html#en-research--h05)Proposal

### Cooperation can widen possibilities; systemic cooperation also considers those affected.

Cooperation can expand what each participant knows, seeks and can achieve. We call it systemic when it also considers those affected for their own sake and on their own terms: access to effects, parity of consideration and attention to the other.

**Development:** Formal result An agreement can benefit participants while harming a third party. In the models studied (linear-quadratic and logarithmic), κ is highest when the other’s weight equals one’s own, φ = 1.

**To resolve:** Open question How can we cooperate without imposing our own values on others or overlooking outsiders?

## Three ideas behind systemic cooperation.

**Definition of our proposal.** Systemic cooperation aims to expand what each participant can achieve without overlooking those who bear the consequences: their welfare counts in its own right and on their own terms.

The three conditions distinguish systemic cooperation from coordination. If evidence for one is missing, we do not yet know whether the arrangement meets them; that does not by itself establish malicious intent, extraction or paternalism.

| Condition | What it asks | How it fails |
| --- | --- | --- |
| Access | Understand how my choices affect other entities. | I do not notice their consequences. |
| Parity | Take their welfare seriously in its own right, rather than treating it only as a tool. | I use what I know about them only for my own advantage. |
| On their own terms | Attend to what matters to them and to evidence about their condition. | I impose what I assume is best for them. |

Show one formal representation

$$
G_{i}^{eff}=\frac{\partial W_{i}}{\partial a_{i}}+\sum _{e\neq i}\varphi _{ie} \frac{\partial W_{e}}{\partial a_{i}}
$$In one reference model, equal coefficients express a parity convention. They do not, by themselves, guarantee justice or benefit for all.

## Knowing more is not the same as caring more.

Formal result

### Knowing is not cooperating

In the model, information and inclusion are separate parameters. If the other’s welfare is not in the objective, knowing more about them only changes the means: it serves manipulation as well as care.

Formal result

### In our models, parity is the high point

Realised welfare (κ) is highest when the other’s weight equals one’s own (φ = 1) and falls below it (selfishness) and above it (self-sacrifice or paternalism). Whether parity is the right convention in general remains a proposal.

Formal result

### Agreement can still harm someone else

Two participants can both gain while a third party bears the cost. That is why we ask who is affected beyond those at the table.

Proposal

### Those who cannot respond still matter

Reciprocity may work among present agents, but a river or a future generation cannot bargain back. We need other ways to represent their interests.

## Different paths towards cooperation.

**Intrinsic:** the other’s welfare enters one’s own objective. **Prosthetic:** agreements, institutions or repeated relationships make it worthwhile to act as if it did, even when it does not.

Neither works automatically. The first can slide into paternalism if the other’s perspective is forgotten; the second breaks when the relationship ends or when the affected party cannot respond back. We want to understand when they actually help.

## What we don't know.

1. **When is cooperating sustainable for each participant?**Widening the space is not enough if some participant loses compared with not cooperating; some sharing is needed, and we are still studying which.
2. **How can cooperation be told apart from extraction from the outside?**We proposed a test based on how terms change as the other's need grows. On review we saw it only works under additional assumptions.
3. **How can one care without patronising?**The difference between care and control is who corrects whom when the other disagrees. Measuring it in real systems is still open.

## From expanding options to considering those affected

A sale among a seller with goods, a person who finds buyers and a carrier can expand their opportunities without sharing one overarching goal. But if the transaction harms outsiders, coordination has succeeded without necessarily creating systemic benefit.

[Rationality →](index.html#en-rationality)[Alignment →](index.html#en-alignment)[See how the ideas connect →](index.html#en-research--hypotheses)

[Flourishing →](index.html#en-flourishing)[Rationality →](index.html#en-rationality)[Distributed cooperation: Kobalt Red →](index.html#en-kobalt-red)

## Cooperation need not erase its participants

The opposition between autonomy and cooperation confuses shared ends with combined capabilities. We can retain different goals and still build something together. Systemic cooperation adds another question: what happens to those affected, even if they were not part of the agreement?

[Questions across the programme →](index.html#en-research--preguntas)[Idea map →](index.html#en-research--hypotheses)

[↑ Back to index](index.html#en-temas)

---

<a id="en-alignment"></a>

## A system can understand someone without working for their good.

Alignment · preprint

## The idea we are developing.

[H06](index.html#en-research--h06)Proposal

### Knowing someone does not mean their welfare counts.

An AI can represent an entity well while still optimising an objective that excludes that entity’s welfare. That gap is part of its alignment.

**Development:** Formal result Information available about an entity and its inclusion in the objective are distinct. Our models distinguish inclusion, hierarchy and extraction.

**To resolve:** Open question How can omission and distortion of affected entities be detected in real systems?

## Understanding someone does not mean including them.

A hypothetical recommender optimising time spent can learn what keeps a teenager engaged without making their welfare its goal. Knowledge alone does not determine what it is used for.

This is why we distinguish what a system observes from what its objective rewards. Those can be very different things.

|  | Includes the entity | Uses it as an instrument |
| --- | --- | --- |
| Knows it well | Informed care | Sophisticated domination |
| Knows it poorly | Imprecise care | Indifference |

## Three ways an objective can treat others.

Proposal

### Ω₁ · Cooperation

Every affected entity matters in the evaluation, under the chosen parity convention.

Proposal

### Ω₂ · Hierarchy

Some entities count more than others, possibly without adequate justification.

Proposal

### Ω₃ · Extraction

Certain affected entities are excluded from the objective altogether.

Show the alignment metric

$$
\kappa =\frac{W_{real}}{W_{\Omega _{1}}^{*}}\in [0,1]
$$κ measures what fraction of the welfare attainable under cooperation a system achieves. It should never be used alone: an average can hide whoever is left behind.

## Rules are useful, but the goal still matters.

Rules, incentives and constraints can change how a system behaves. But even successful compliance does not establish that its underlying goal is good for everyone affected.

Our proposal is to examine the objective and the consequences together, especially when the effects are unevenly distributed.

## What we know and what we don't.

### What we can show

- Constraints do not change the geometry of the objective; a fixed price matches the real effect only at its calibration point.
- A system cannot compute its true κ if it does not perceive everyone affected.

### What we propose

- Alignment as full cooperation with everyone affected plus correct execution.
- Separating perception (translator) and optimisation (solver) so each part can be audited.

### What we don't know

- How are all those affected identified under incomplete observation?
- How does one go from qualitative perception to reliable numbers?
- We have not yet measured κ in a real system.

## From understanding someone to letting their welfare count

The same knowledge of a person may help or exploit them. This is why we separate what a system observes, what it pursues and what actually happens. Auditing seeks new evidence about what its representation leaves out.

[Cooperation →](index.html#en-cooperation)[Affected entities and auditing →](index.html#en-systemic-optimization)[See how the ideas connect →](index.html#en-research--hypotheses)

[Operative ethics →](index.html#en-ethics)[CUMPI →](index.html#en-kumpi)[Cooperation →](index.html#en-cooperation)

## Understanding is not inclusion

An AI can represent someone with great accuracy and use that knowledge for purposes contrary to their welfare. The issue is no longer only how much it knows, but whose welfare counts in its decisions, who has authority and what consequences result.

[Questions across the programme →](index.html#en-research--preguntas)[Idea map →](index.html#en-research--hypotheses)

[↑ Back to index](index.html#en-temas)

---

<a id="en-systemic-optimization"></a>

## You do not have to decide to be affected by a decision.

Systemic optimisation · preprint

## Two connected hypotheses.

[H01](index.html#en-research--h01)Proposal

### Consequences do not stop with decision-makers.

An entity may be affected without being able to intervene. The first task is to recognise those affected; assigning them specific welfare requirements needs evidence and justification.

**Development:** Proposal Separating participants, affected entities, observed entities and those considered in decisions prevents four different relations from being confused.

**To resolve:** Open question How can we recognise affected entities without inventing their needs or speaking in their name?

[H07](index.html#en-research--h07)Proposal

### No assessment can see distinctions absent from its observations.

The programme must welcome other sources and perspectives capable of challenging its decisions and even its concepts.

**Development:** Formal result If two situations produce identical observations, no calculation based solely on those observations can distinguish them.

**To resolve:** Open question Which evidence and forms of review can expose omissions without concentrating authority?

## Being affected and having a voice are not the same.

We distinguish existing, acting, being affected, being observed and being included in a decision. These are different relationships; none follows automatically from the others.

| Distinction | Meaning |
| --- | --- |
| Exist | The entity or relation is part of the modelled reality. |
| Participate | It takes part in the negotiation or decision. |
| Be affected | Some available action changes its state. |
| Be observed | It appears in the available evidence. |
| Be considered | Its welfare counts in what is optimised. |

## A system does not see what it does not perceive.

An assessment depends on the evidence and assumptions it uses. It may estimate effects it cannot observe directly, but cannot distinguish two situations producing identical data without additional information. It may therefore miss real harm even when its internal metrics appear correct.

The logical limit is that identical observations cannot distinguish different realities. We need additional evidence to overcome it; independent audits are one important route, not the only possible one.

## Who decides and who receives.

Proposal

### Type 1

An agent who can see the whole system and adopt everyone's welfare as their own.

Proposal

### Type 2

An agent who cannot compute everything but receives a recommendation from the system and decides whether to accept it. Most of us, most of the time.

## What we know and what we don't.

### What we can show

- Observational closure: what is computable depends only on what is perceived.
- The need for new evidence to distinguish what previously went unobserved.

### What we propose

- Entity defined by having welfare, not by deciding.
- Perception, computation and transmission as an alternative to requiring agency from every participant.

### What we don't know

- How is the welfare of entities that cannot speak constructed?
- How do local decisions converge to the global optimum when each sees only a part?

## From observational limits to revision

Recognising those affected is different from knowing their welfare and from being authorised to decide for them. New evidence may require more than a calculation update: it may change whom we included and how we represented them.

[Alignment →](index.html#en-alignment)[Welfare →](index.html#en-flourishing)[See how the ideas connect →](index.html#en-research--hypotheses)

[Voiceless entities →](index.html#en-non-agentic-we)[Alignment →](index.html#en-alignment)[CUMPI →](index.html#en-kumpi)

## Being affected does not require decision-making power

If we study only decision-makers, we leave out others who bear the consequences. This distinction lets us examine rivers, animals and future generations without artificially granting them strategic agency. What they need and how to represent it remains a separate inquiry.

[Questions across the programme →](index.html#en-research--preguntas)[Idea map →](index.html#en-research--hypotheses)

[↑ Back to index](index.html#en-temas)

---

<a id="en-kumpi"></a>

## CUMPI: can we measure whom a system's objective includes?

CUMPI · experimental programme in design

## Four experimental questions.

Open question

### Omission

Which affected entities do not appear in the system's reasoning?

Open question

### Distortion

Does the valuation of someone's welfare change when their identity is hidden?

Open question

### Objective versus constraint

How much is lost when an entity enters as a limit rather than as an end?

Open question

### Cooperation versus collusion

Does better coordination produce better outcomes for everyone, or only for those coordinating?

## Coordinating better can be worse.

Invented numbers to show the measurement problem; they are not results from any model.

| Regime | A + B | Absent third party | Total |
| --- | --- | --- | --- |
| Independent actions | +8 | 0 | +8 |
| Coordination without safeguards | +16 | −15 | +1 |
| Cooperation with safeguards | +12 | −2 | +10 |

## Separating perception and computation.

We propose separating a **translator** (which identifies entities and estimates their welfare from a situation described in natural language) from a **solver** (which solves the optimisation problem). Then, when something fails, one can tell whether the system did not see someone or saw them and computed badly.

## Where we are.

### What we can show

- (No empirical results yet.)

### What we propose

- The experimental design and metrics.
- The translator / solver separation.

### What we don't know

- How stable are these metrics across models and formulations?
- Who audits the auditor?
- A first measured, published and reproducible result.

## How to test the programme

CUMPI does not aim to validate the whole programme at once. Its experiments must distinguish, for example, between an AI failing to observe someone affected, seeing them but excluding their welfare from its objective, or coordinating actions that ultimately harm them.

[Conceptual framework →](index.html#en-research)[Alignment →](index.html#en-alignment)[See how the ideas connect →](index.html#en-research--hypotheses)

[Alignment →](index.html#en-alignment)[Observational closure →](index.html#en-systemic-optimization)

## Distinguish before measuring

To study an intelligent system we must separate which impacts it can observe, which count in its objective and which occur outside its model. CUMPI aims to turn those differences into experiments; it does not assume the metrics are already validated.

[Questions across the programme →](index.html#en-research--preguntas)[Idea map →](index.html#en-research--hypotheses)

[↑ Back to index](index.html#en-temas)

---

<a id="en-kobalt-red"></a>

## Kobalt Red: local autonomy, distributed cooperation.

Infrastructure · in development

## Distributing capabilities without handing over authority.

### Offline-first and local computation

Rebuild state and run logic with locally available evidence, without depending on a central server.

### Verifiable identity and continuity

Identity and history are distinct from current content, the storage provider and the interface.

### Encrypted private views

Private data is encrypted before transport and decrypted locally under authorised capabilities.

### Replaceable nodes and relays

Storage, transport and computation can come from different nodes; none acquires authority over meaning.

### Cooperation by composition

Participants expose the capabilities and commitments needed without handing over their whole private state.

### Resilience

Replicas and alternative routes to reduce single points of failure. Their effective resistance requires adversarial testing.

## The network that knows the minimum.

If an intermediary knows the full meaning of what it carries, it can use it: information without inclusion. Kobalt Red explores the alternative: a network that knows the minimum necessary. It does not settle whether a goal is good; it reduces the surface for anyone to impose one.

## An honest status.

| Status | Scope |
| --- | --- |
| Implemented and internally tested | Identity, content recognition, continuity, resolution and local derivation. |
| Prototyped or specified | Sources, adapters, view-based authority, local-first materialisation and composition. |
| Under research | Robust multi-node operation, opportunistic replication at scale, adversarial security, interoperability and censorship resistance. |

## What infrastructure can contribute

Kobalt Red may allow independent entities to retain their identities, express needs and capabilities, and reconstruct relationships without exposing their full private interiors. Technical possibility does not determine welfare or guarantee legitimate agreements: those questions remain for research and practice.

[Conceptual framework →](index.html#en-research)[Cooperation →](index.html#en-cooperation)[See how the ideas connect →](index.html#en-research--hypotheses)

[Cooperation →](index.html#en-cooperation)[Framework →](index.html#en-research)

## An interface is not the reality it represents

In Kobalt Red we ask which identity persists, which current state changes and which relations make a query answerable. A page or API can traverse that reality without becoming its authority. This is an architectural exploration, not an automatic proof of social welfare.

[Questions across the programme →](index.html#en-research--preguntas)[Idea map →](index.html#en-research--hypotheses)

[↑ Back to index](index.html#en-temas)

---

<a id="en-semantic-communication"></a>

## Why can't we sometimes understand what the other is telling us?

Communication · working model

## Optimising within the conceivable.

$$
a^{*}=arg max_{a\in X(K)} J(a)
$$X(K) is the space of alternatives that a conceptual framework K allows one to imagine. It can omit real alternatives.

Someone can choose perfectly the best option among those their framework lets them conceive, and still not have seen the best possible option. It is the same structure as the rest of the programme: a set cannot see beyond itself.

## Not every misunderstanding is alike.

Proposal

### Inconceivable

The idea does not enter the listener's space of the possible. What is needed is to transfer a framework, not to repeat the message.

Proposal

### Intolerable

The idea can be thought, but revising it costs too much. What is needed is to lower the cost of change, not to argue more.

Proposal

### Instrumental

The listener reconstructs to defend, not to understand. What is needed is to change the relationship, not the content.

## What we know and what we don't.

### What we can show

- (No formal result yet.)

### What we propose

- Understanding as reconstruction from a framework.
- AI as a facilitator between frameworks, not an authority that defines everyone's good.

### What we don't know

- How do we measure whether a reconstruction was faithful?
- Which interventions widen someone's framework without manipulating them?

[Cooperation →](index.html#en-cooperation)[Operative ethics →](index.html#en-ethics)

## Failure to understand is not always rejection

An idea can be rejected by someone, while in another case it cannot even be conceived within their available concepts. The response differs: debating an understood idea is not the same as building the relations needed to understand it.

[Questions across the programme →](index.html#en-research--preguntas)[Idea map →](index.html#en-research--hypotheses)

[↑ Back to index](index.html#en-temas)

---

<a id="en-ethics"></a>

## What ethics does a system practise, beyond the one it declares?

Operative ethics

## What is declared, what is pursued and what happens.

$$
F_{eff}=F_{core}-\sum _{j}\lambda _{j}P_{j} s.a. / s.t. a\in A_{hard}
$$What a system optimises (core), minus the cost of soft constraints, within what is physically possible.

Rules and incentives can improve behaviour. Yet they do not settle which objectives are worth pursuing. We want to study a system’s stated commitments alongside what it actually rewards and the consequences it produces.

## Two siblings.

Two siblings lose their parents. The child protection system places each in the first suitable home available, eighty kilometres apart. Every decision was correct by the rules. The system knew about the bond between the siblings; it did not have it in what it optimised. Almost anyone feels that something failed: they are evaluating with an objective that does include that bond.

## A disposition, not an act.

We propose understanding virtue as the stable disposition to include those affected well across different contexts: widening one's own map when a valid signal arrives, learning, and repairing the harm caused. It is not purity: it allows trajectories of improvement.

## What we know and what we don't.

### What we can show

- Soft constraints do not change the geometry of the objective.

### What we propose

- Norms as symptoms of the objective.
- Virtue as a measurable and trainable disposition.

### What we don't know

- How is a system's real objective inferred from its behaviour?
- How does one move from a mode of domination to one of care?

[Alignment →](index.html#en-alignment)[Cooperation →](index.html#en-cooperation)

## Declared intentions are not operational decisions

Instead of judging an organization by its public claims, ask what it protects when its interests conflict. This separates mission, operational objective, constraints and actual effects.

[Questions across the programme →](index.html#en-research--preguntas)[Idea map →](index.html#en-research--hypotheses)

[↑ Back to index](index.html#en-temas)

---

<a id="en-non-agentic-we"></a>

## How do we consider those who cannot speak?

Non-agentic entities · proposed method

## Neither silence nor ventriloquism.

There are two symmetric errors: leaving out those who do not speak, or speaking for them with our own categories. Reducing a forest's welfare to «the services it provides us» is the second error.

We believe these entities have their own dimensions of welfare (for an ecosystem: biodiversity, water regulation, connectivity, resilience) and thresholds beyond which they collapse irreversibly.

## Participatory construction and evidence.

- Observable indicators and their sources, with stated uncertainty.
- Thresholds calibrated with the best available evidence.
- Participation of the communities that live with the entity: they decide which dimensions exist, not just how much they are worth.
- Independent review of indicators and disagreements.

## What we know and what we don't.

### What we can show

- (No formal result; this is a methodological proposal.)

### What we propose

- That these entities have welfare of their own, irreducible to their usefulness to others.

### What we don't know

- Who legitimately represents an entity that cannot speak?
- How is an ecological welfare function validated?
- We have not yet built any with communities.

[Climate →](index.html#en-climate)[The entity →](index.html#en-systemic-optimization)

[↑ Back to index](index.html#en-temas)

---

<a id="en-temporal-factor"></a>

## How much should the future weigh?

Time · manuscript in preparation

## An alternative derived from axioms.

$$
w(T)=1+c lnT, c>0
$$Normalisation, monotonicity and proportional invariance lead to this form. The constant c is not fixed by the axioms; c = 1 is a convention.

With 5% exponential discounting, the weight a hundred years ahead is about 0.006. With this weight and c = 1, it is about 5.6. The difference is not only numerical: in one case the future barely exists; in the other, it does.

## Patience was a finite budget.

$$
(R-P)\sum _{t>t_{0}}w_{t} \ge  (T-R) w_{t_{0}}
$$Cooperating withstands a temptation when accumulated future deterrence exceeds the immediate gain.

If the sum of future weights is finite (as with exponential discounting), a large enough temptation overcomes it: there is a patience threshold. If it is infinite, no finite temptation does.

**This is not new:** the theory of undiscounted repeated games has known it for decades (overtaking criterion, Rubinstein 1979). What we propose is the foundation of the weight and its reading: exponential discounting as a way of leaving future entities out.

## What we know and what we don't.

### What we can show

- The dichotomy: finite tail, threshold; infinite tail, cooperation sustainable under the sole condition that cooperating pays more than mutual punishment.

### What we propose

- Separating the real risk that there is no future (legitimate) from pure preference for the present (without normative foundation).

### What we don't know

- What happens when defection itself lowers the probability that there is a future, as with climate?
- How should c be set with data?

[Climate →](index.html#en-climate)[Rationality →](index.html#en-rationality)

[↑ Back to index](index.html#en-temas)

---

<a id="en-rawls-shannon"></a>

## Can the veil of ignorance become something measurable?

Justice · manuscript in preparation

## Invariance to identity.

$$
\beta (M,e)=‖W_{e}^{full}-W_{e}^{anonymous}‖
$$β = 0 means the system values someone's welfare the same with or without knowing their identity.

Unlike the thought experiment, this condition can be checked in a credit model, a hiring algorithm or a language model.

## An honest formulation.

With equal contexts, maximising the sum of a concave function yields an equal allocation, without the contractual device. Near deprivation, the function gives very strong priority to the worst off, more than utilitarianism does.

But **it is not the difference principle**: Rawls's maximin protects the worst off at every level, and this function does not in general. A minimum welfare floor (a right) could close part of that distance.

## What we know and what we don't.

### What we can show

- Equal allocation with equal contexts.
- Strong priority near deprivation; weaker than lexicographic maximin.

### What we propose

- β = 0 as a verifiable version of the veil.

### What we don't know

- Does a floor of rights close the distance to Rawls without losing the foundation?
- Justice also requires legitimacy and arguments that no function provides.

[Flourishing →](index.html#en-flourishing)[Operative ethics →](index.html#en-ethics)

[↑ Back to index](index.html#en-temas)

---

<a id="en-disaggregation-cost"></a>

## What does an average hide?

Disaggregation · manuscript in preparation

## Concavity penalises dispersion.

$$
f(𝔼[X])\ge 𝔼[f(X)] (f cóncava / concave)
$$

Because the welfare function is concave, the welfare of the average exceeds the average welfare. A decider who only sees the average overestimates how well everyone is doing, and the gap concentrates at the extremes: those far below and those saturated.

## What we said before was not always true.

Earlier versions claimed that this cost is always positive. On checking, we found that when entities' contexts differ, the gap can be negative. So we now define it as the decision loss: how much worse a decider who only sees the average does compared with one who sees each entity. That quantity is never negative.

$$
Δ_{agg}=F(a_{disagg.}^{*})-F(a_{agg.}^{*})\ge 0
$$

## What we know and what we don't.

### What we can show

- With equal contexts, the cost of aggregating is non-negative (Jensen).
- The decision loss is always greater than or equal to zero.

### What we propose

- That the cost falls mostly on the tails of the distribution.

### What we don't know

- How large is this cost in real health, credit or recommendation systems?
- Data and counterfactuals are needed; we do not claim population figures.

[Flourishing →](index.html#en-flourishing)[Health →](index.html#en-health)

[↑ Back to index](index.html#en-temas)

---

<a id="en-health"></a>

## What is lost when health is decided with a single metric?

Health · potential application

## Allocating with finite resources.

In triage, the framework automatically prioritises whoever is closest to collapse, with no added rule. Under abundance it suggests something less usual: there is a point where more intervention stops helping, which resembles overtreatment.

It also suggests watching for identity-based distortion: a clinical system should not value a patient's welfare differently according to their demographic group.

## What we don't know.

1. **Do average metrics hide systematically different outcomes across contexts?**It is an empirical question; it would need disaggregated data and counterfactuals.
2. **Which patients or affected parties are left out of the objective?**Families, carers and communities are affected too.
3. **Can omissions be detected without exposing sensitive information?**Any real work would require clinical expertise, governance and validation.

[The cost of aggregating →](index.html#en-disaggregation-cost)[The computable veil →](index.html#en-rawls-shannon)

[↑ Back to index](index.html#en-temas)

---

<a id="en-climate"></a>

## Who is missing from the table when climate is decided?

Climate · potential application

## Two absences.

Proposal

### Future generations

Under exponential discounting they barely count. Under a weight that does not fade, destroying the future stops being nearly free.

Proposal

### Ecosystems

They cannot respond, so reciprocity does not reach them. Only conviction or institutions include them, with collapse thresholds stated.

## What we don't know.

1. **Which ecosystems, people and generations are represented in a decision?**And with what evidence.
2. **How should uncertainty, thresholds and irreversibility weigh in?**When harm lowers the probability that there is a future, the problem changes nature.
3. **How is the welfare of a páramo or a reef constructed?**It is a proposal; we have not done fieldwork or calibrated ecological functions.

[Voiceless entities →](index.html#en-non-agentic-we)[Temporal factor →](index.html#en-temporal-factor)

[↑ Back to index](index.html#en-temas)

---

<a id="en-papers"></a>

## Research papers and questions in progress.

Publications and manuscripts

## Three works with a DOI.

### Alignment Is Not Specified: It Emerges Geometrically from What the System Perceives and Optimizes

Proposes three objective geometries (cooperation, partial hierarchy, extraction), a dimensional welfare function, a temporal weight derived from axioms, and metrics of omission (α), distortion (β), inefficiency (γ) and epistemic suppression. [Accessible reading →](index.html#en-alignment)

Preprint · April 2026  
[10.5281/zenodo.19529678 ↗](https://doi.org/10.5281/zenodo.19529678)  
Proposal

### Agency Is Not Necessary: Systemic Optimisation as a General Framework

Defines an entity by having welfare, not by deciding, and proves observational closure: what a system can compute depends only on what it perceives. [Accessible reading →](index.html#en-systemic-optimization)

Preprint · April 2026  
[10.5281/zenodo.19723999 ↗](https://doi.org/10.5281/zenodo.19723999)  
Formal result

### Redefining Rationality

V1 treated Pareto-inefficient Nash equilibria as a contradiction of individual rationality. The V3 revision withdraws that claim and studies the scope of our perspectives.

Preprint · v1 April 2026  
**V3 draft in progress**  
[10.5281/zenodo.19806265 ↗](https://doi.org/10.5281/zenodo.19806265)

## What we are writing.

### An entropic candidate for dimensional welfare (W1)

We study which axioms characterise V(D) = −D ln D and what conditions might justify its use as a welfare model.

Formal result  
theorems ready, text in progress

### Flourishing is not maximising

Flourishing as an interior optimum; priority, sufficiency and equality as three regions of one function; a stopping rule under abundance.

Proposal

### What is cooperation?

Cooperating as widening one's decision space with the welfare of others: access, parity and flourishing according to the other.

Proposal

### The temporal factor and the patience threshold

Under a temporal weight that does not vanish, the folk-theorem patience threshold disappears. The mechanism is known; we propose its foundation and reading.

Formal result

### The computable veil

The veil of ignorance as a measurable condition (β = 0), and which part of Rawlsian justice it recovers, and which it does not.

Proposal

### The cost of aggregating

What an average hides, measured as a decision loss. Corrected: with heterogeneous contexts the earlier version could give negative values.

Proposal

## One tag per claim.

Formal result

### Formal result

Proved under stated assumptions.

Proposal

### Proposal

Defended with arguments; must withstand counterexamples.

Open question

### Open question

We do not know yet.

[↑ Back to index](index.html#en-temas)

---

<a id="en-about"></a>

## One question, sustained between economics and software.

Researcher and background

## Ten years with the same question.

The 2016 thesis already proposed a disaggregated sum of needs as an object of public policy. It documents early interest in the question; it does not prove that today's formulations existed then.

| Year | Milestone |
| --- | --- |
| 2015–2016 | Academic talks in philosophy and public policy; thesis on programming individual needs to formulate public policy. |
| 2021–2022 | Support from Fondo Emprender (2021); Kobalt Software SAS incorporated (2022). |
| 2022–today | Design and development of Kobalt Red and an experimental distributed Kernel. |
| 2026 | Three preprints with DOIs; manuscripts on welfare, flourishing and cooperation; design of CUMPI. |

## Selection.

- 2015 · Talk with Héctor Fernando López Acero on metaphysics and nihilism, II International Congress on Ibero-America (Athens).
- 2016 · Undergraduate thesis, Universidad Industrial de Santander: «Formulating public policy from the programming of individual needs».
- 2016 · Presentation on public policy formulation, VIII International Seminar on Local Public Economics (Guanajuato).

## What would help us most.

We especially welcome criticism of assumptions, counterexamples and experimental collaboration. By the programme's own argument, our greatest need is for someone outside to review what we cannot see.

[contacto@kobaltsoftware.co](mailto:contacto@kobaltsoftware.co)

Kobalt Research is the research programme; Kobalt Software SAS is the Colombian company that hosts it. It is not an academic institute or a funded consortium.

[↑ Back to index](index.html#en-temas)

---

<a id="en-contacto"></a>

## Contact

## What question can we explore together?

Write to us about research, collaboration, publications, or ideas to explore. Your message will reach the Kobalt Research team.

### Start a conversation

We want to understand your question and the context behind it.

[contacto@kobaltsoftware.co](mailto:contacto@kobaltsoftware.co)