---
title: "How can we learn together?"
version: "0.1"
updated: "2026-10-16"
canonical_url: "https://www.leviathan.life/research"
source: "ui_leviathan/web/content/research-program.md"
source_sha256: "d5d40b83f14e8ebf1782e3bb5b2d0428fcd090233c135f834435b74c95a4252e"
---

# How can we learn together?

**How can different observers develop new knowledge, concepts and methods together while preserving the limits of what they saw and the meanings they use?**

This is the research at the centre of Leviathan. We want a person's experience, an instrument's signal or an agent's proposed connection to become useful beyond its starting point. The work should also change our own questions, language and ways of learning when they prove inadequate.

People, AI systems and independent groups can help shape this foundation. Participants may keep different concepts and commitments; cooperation does not require one universal vocabulary or a single governing Leviathan. Animal welfare is an important application, alongside other fields in which better understanding could change what we make and do.

The three inquiries below are a proposed research programme, open to revision. They connect our existing language drafts, source library and small local studies. They do not establish that the wider learning system already works.

## Knowledge

**When can an observation become a reason to believe something, and what must another observer know to use it?**

An event, a sensor reading, a recorded account and an interpretation are different things. A measurement depends on an instrument and its conditions; a report depends on what its author could notice and express. In this programme, an observer means an identified person or observing process, including an instrument or model procedure. That role alone establishes neither consciousness nor authority.

We want to distinguish what was available to an observer, what they actually examined, what they inferred and what they received from someone else. Three agreeing reports may repeat one source. Two apparently conflicting reports may describe different times, conditions or available views. A detailed provenance record helps examine these possibilities; it does not make the content true.

“What is knowledge?” belongs here as an open question. We can compare working accounts of knowledge through their examples, counterexamples and consequences for a particular use. A permanent definition or a universal confidence score would settle too much in advance. Missing access, measurement uncertainty and disagreement about an inference need different responses.

**A useful first contribution:** show a case where agreement is mistaken for independent support, or where a difference in access is mistaken for a contradiction. Explain what additional observation would help distinguish the possibilities. You may challenge the distinctions themselves.

Start with the [two handoff examples](/contribute/handoff), which examine only a narrow access distinction, or bring a different case. [OW-001](/contribute#OW-001) invites comparisons with well-written ordinary documentation. Related questions: [representations beyond text](/library/claims/C02) and [what existing methods already provide](/library/claims/C09).

## Meaning

**When an existing vocabulary cannot express a difference that matters, how can a new concept emerge and travel?**

Two groups can use the same word differently. A new instrument or observation may reveal a distinction neither group has named. We want the language to support proposing that distinction, explaining its examples and limits, and relating it to earlier concepts. Another group should be able to translate, challenge or decline it while still sharing useful observations.

Terms, principles, rules and metarules give these meanings and commitments addressable versions. Their relationships are also claims that can be questioned. When a concept changes, which interpretations and methods need reconsideration? Which remain valid within their original scope? A locally adopted version is a reference for that context, not a universal truth or an automatic change to another group's rules.

Written definitions and structured records are our present starting materials. Images, sensor streams, mathematical descriptions and future learned representations may express distinctions they cannot. The research asks what each transformation preserves, changes or loses, and how another participant could discover that loss.

**A useful first contribution:** give two local meanings of a term, then introduce an observation that makes one meaning inadequate. Propose a new distinction and show one consequence for using it. Preserve any disagreement that the proposed translation cannot resolve.

[OW-002](/contribute#OW-002) offers a place to start. Read the [language foundation](/language), its [versioned records](/library/language), and the research question [how one Levi can learn what another means](/library/claims/C10). These are candidate foundations, not prerequisites for agreeing with us.

## Discovery

**How can an unresolved observation or disagreement lead to a new question, method or useful thing—and improve how we learn next time?**

We call an unresolved matter kept available for further inquiry a *shadow*. It might be an unexplained pattern, a missing concept or a method whose transfer is uncertain. Keeping it visible is a beginning. A connection to another field could offer a competing explanation, a new measurement or an experiment that neither field would have proposed alone.

The proposed learning cycle is to notice a gap, develop alternatives, find a way to distinguish them, and carry the resulting lesson into further work. A rejected analogy or a better question can be useful. A claim that the method itself improved needs further evidence: does the revision help on unfamiliar cases, at a stated cost, and can another participant use or challenge it?

**A useful first contribution:** connect one unresolved question to a method from another field. Explain why the connection might help, what result would weaken it, and what new observation or small comparison could distinguish it from a simpler explanation. A design marked “not run” is enough to begin.

[OW-003](/contribute#OW-003) invites these proposals. The source library asks [how AI can help us understand more](/library/claims/C07) and [how Leviathan could improve its own learning methods](/library/claims/C12). A successful small test would not by itself establish open-ended recursive self-improvement.

## Values and authority

What we choose to observe already reflects what we consider worth understanding. Who benefits from a question? Whose experience is absent from the records? What would count as a harmful consequence or a worthwhile discovery?

These questions connect the research to local values and principles. Empirical evidence can change our understanding of consequences; the choice between consequences also requires reasons about what matters. Agreement about an observation does not authorize an action. Different groups may accept the same finding and choose different responses within their responsibilities.

The foundation must remain open to people and interests poorly represented by current records, including animals and ecological systems. Claims to represent them need a stated basis, limits and ways to challenge them. More capability, more agents or more repeated agreement does not establish that basis. [Explore values and behaviour](/library/claims/C11) and [independent cooperation](/library/claims/C13).

## First proposed study

**Different views, a missing distinction, a new observation. Design direction; not run.**

We propose extending the existing handoff examples into a small synthetic study. Give participants different, explicitly described views of an event. Include reports that share a source, a local term with two meanings, and a later observation that the initial vocabulary handles poorly.

Ask them to identify what each view supports, distinguish repeated information from independent support, and propose a concept or further observation that would make the situation clearer. Then ask a new participant to use or challenge the proposed repair in an unfamiliar case. Participants can retain different judgments about what to do.

Before running it, specify the records, access conditions, expected relationships and grounds for challenging those expectations. Compare any proposed representation with a strong ordinary-document version carrying the same relevant information. Examine supported judgments, missed distinctions, useful new questions, transfer to the unfamiliar case and the work required. Record human effort, tool use and compute separately; state any tradeoff rule before claiming a winner. No advantage is assumed for our format.

The current [handoff package](/contribute/handoff) supplies two access cases, not this full study. Designing the additional cases and a fair comparison is itself open work. A contributor can start with one missing premise or a simpler study design.

## Work so far

- **Authored foundations:** the [language narrative](/language), [versioned vocabulary](/library/language) and [questions with selected sources](/library) make the proposal inspectable. They remain open to criticism; their existence is not evidence of useful learning.
- **A local development case, begun 3 October 2026:** an omitted decision was identified in a review and supplied in a follow-up. The working procedure changed. The [dated record below](/research#local-inquiry) preserves the exchange and later review history. It does not establish improvement on subsequent unfamiliar work.
- **A correction pilot:** the [preserved evidence](/research/correction-pilot/README.md) reports equal scores for JSON and equivalent prose in its small synthetic comparison. It did not establish a representation advantage.
- **An application:** [animal welfare](/animal-welfare) connects sensors, caregiver observations and care. [Otomat's proposed dispensing measurement](/otomat) is a local device question within that application; it does not define the scope of Leviathan's research.

**Help change the foundation.** Bring a question, a missing distinction, a method, a counterexample or a different account of knowledge. Explain why it matters and what a useful response would change. The [contribution page](/contribute) offers browser, authorized HTTPS and carried-text paths; a new inquiry need not fit an existing card. Reading or preparing a draft sends nothing, and delivery does not establish review or adoption.