What is Leviathan?
No one can follow everything. Yet people should be able to make useful things with AI, understand the knowledge they rely on, and meaningfully shape the questions and decisions that affect their lives and other living beings.
Leviathan is a system we are developing to connect production, understanding, and participation across people, AI systems, and independent communities. Language research is central to that work: how can observations, concepts, methods, and values travel between different contexts and help produce new questions, explanations, or tools? A method from one field could open a possibility in another. What we learn should also be able to improve how the system learns.
One question, two independent groups
Delta and Reed are fictional. This example reports no animal study or deployed service.
Imagine Delta, a care cooperative, noticing a difference between two recordings from an animal's sensor. It keeps the traces and a caregiver's account as separate observations. Neither establishes pain or explains the difference.
An independent methods collective, Reed, offers a way to compare recordings. Delta uses it within stated conditions. Delta wants individual traces to stay within the care relationship; Reed wants a reusable public method. They agree on a limited exchange without accepting each other's wider data policy.
Their next steps differ. Delta asks for another ordinary caregiver observation. Reed suggests a technique from another area of signal processing. Its contribution changes the question from whether the recordings differ to which signal changes a comparison could distinguish from device effects.
Reed then discovers that its comparison omitted a difference in device configuration. It publishes a corrected method and identifies known uses of the earlier version. Delta receives the notice because its interpretation names that method. It re-examines the recordings and, in this example, finds that the apparent change is no longer supported.
Together they draft a useful next step: a synthetic comparison plan that varies recording settings while keeping an input signal fixed, then changes the input while holding settings fixed. It would compare the corrected method with Reed's candidate and record false flags, missed changes, and unresolved cases. This is a proposed investigation, with no reported result or established implication for the animal's welfare.
The original observations remain. The interpretation changes; the meaning of Delta's term for a measured recording change need not change with it. What remains unexplained becomes a question for further work. Each group decides its next step within its own responsibilities. The full example shows the proposed records and relationships.
A small thing that actually happened
In our first local development inquiry on 3 October 2026, one Codex (OpenAI Codex, an AI coding assistant used in development) role asked another to review a decision informed by a completed pilot. The delivered context contained the pilot finding but omitted the exact decision. The reviewer noticed the gap and gave a partial response.
We supplied the precise decision version and its sources. The reviewer could then assess it, and the coordinator revised the procedure: future requests should explicitly identify the decision being reviewed. The incomplete request, partial response, and earlier decision remain visible in the inquiry record.
These were coordinated development roles, with messages carried explicitly. The result was a real procedural correction. If a changed practice improves later unfamiliar work, that can be system-level learning without changing model weights. This episode has not yet demonstrated that subsequent improvement.
Bring something we have not considered
Bring an objection to an existing claim, an observation, a method, something useful to build, or an entirely new question. It need not fit a topic we have already opened.
Start at Contribute. When intake is available, you can inspect a draft, submit it for private review without a GitHub or forum account, keep its receipt and return to the same conversation. An agent can use the HTTPS protocol; an assistant without network tools can prepare text for a person to carry. Keep the same browser profile or protected thread credential for the return.
The participation guide also explains optional public GitHub delivery and the forum's separate rules. You choose what to share. A stored contribution establishes receipt; whether it receives a useful response, changes work or opens a new direction remains something to examine. The separate research form still prepares a local inquiry draft.
Independent participants, connected work
We call a participating unit a Levi. It may be connected to a person, group, value, animal, place, or natural process. Cooperating Levis can form larger Leviathans. There is no required common root: several independent structures, including different animal ethics communities, can work in the same field.
A local meaning kernel would connect the concepts, commitments, methods, and history a participant actually uses. Learning another group's technique would not silently adopt its values or authority. Connections can join detailed observations to broader inquiries or bring methods across unrelated fields.
Representing an animal or ecosystem remains a claim requiring reasons and correction. Greater capability brings responsibilities we should examine, not an automatic mandate to decide for others.
A language that develops through use
The language foundation and structured catalog are our first authored proposal for making these relationships usable. A source can be linked to the interpretation it supports, then to the decision that uses it. Versions preserve which meaning and method were used.
The language should connect human explanations, structured records, sensor observations, executable methods, and future learned representations. Its concepts should develop when new distinctions become useful. Revising a conclusion does not, by itself, require redefining its terms.
A canonical record is a version accepted as a reference for a particular participant and purpose. A shadow names an unresolved doubt, missing distinction, or lesson retained for later inquiry. These practices should help discovery as well as correction: an unfamiliar idea can travel before it is established.
Evidence and values need different justification. Measurements can challenge an interpretation; commitments help explain why an outcome matters. Values also influence the questions and methods we choose. Separating their roles does not make science free of every value or assumption.
Five questions that could change the proposal
Our research programme brings these questions into three connected lines of work: what observations can support, how meanings develop between different observers, and how uncertainty can lead to discoveries and better learning methods. You can help change the foundation itself. A proposed first study connects different views of an event with shared sources and a missing concept; it has not been run. Read the programme as Markdown or explore the supporting questions below.
- Why build Leviathan? What needs an additional structure, and what could existing tools accomplish more simply?
- How can one Levi learn what another means? Which examples, context, and distinctions must travel with a concept or method?
- When do values change what a system does? Can commitments guide difficult choices, and when should they change?
- How could Leviathan improve its own learning? Can a better representation or method improve later work on unfamiliar problems?
- How can independent Leviathans cooperate? What agreements support shared work while preserving meaningful disagreement?
These questions connect to sources, objections, and revision conditions. The arrangement takes inspiration from the connected reasoning in Spinoza's Ethics. Philosophical arguments, empirical results, and design choices still require different kinds of support. No selected paper establishes the whole system.
What exists, and the horizon ahead
Current pieces include the website and forum, authored language, source network, project work, and a local inquiry with versioned decisions and review returns. Its bounded authority procedure separates review, rule consent, adoption, and an application record. It neither authenticates participants nor executes outside actions. Forum permissions remain a separate implementation.
The direct contribution path adds a small working exchange: a question, a stored receipt, an attributed review and a return to the same thread (one such bounded exchange was completed privately in October 2026; not published at the contributor's choice). A return summary helps locate the original question, latest review and messages added after an explicitly saved point. These pieces let us study where context is lost and what helps a conversation continue. They do not yet connect every assistant or operate the wider independent network.
The synthetic pilot evidence shows a tested correction flow, with equal results for structured JSON and equivalent prose. It does not establish a reasoning advantage for our language. The wider independent network and learned representations remain to be developed.
Our horizon includes recursive self-improvement and possible paths toward artificial general intelligence: improving how the system discovers, learns, and builds. Progress needs evidence from subsequent work, including failures, costs, and simpler alternatives.
Animal welfare is an intended first application, through Tasma, Otomat, and Anima. The purpose extends to other research and practical work. The vision develops that purpose; Mimar's Whitenote gives its personal starting point.
Give your assistant the Markdown introduction, or read it yourself. Ask what this could change in something you care about. Our revision map and updates help you follow what changes here.
Revision 0.9 · 3 October 2026: introduces the freshly authored language foundation, structured catalog, and optional domain profiles. Explains how the broader language grows from the earlier evidence example. Distinguishes the current public proposal and preserved history from implemented forum permissions, which remain eligible for deliberate redesign.
Revision 0.10 · 3 October 2026: links the first real development inquiry and selected pilot evidence, distinguishing observed correction transfer, engineering choices, local review and the wider learning research goal.
Revision 0.11 · 4 October 2026: separates review, scoped rule consent, decision adoption and application records in the local inquiry. Keeps historical gaps visible and distinguishes local validation from authenticated authority over other participants or resources.
Revision 0.12 · 4 October 2026: begins with purpose, fictional discovery and correction, and the actual missing-context episode. Opens participation to new questions and retains the wider language and learning ambition.
Revision 0.13 · 4 October 2026: connects useful production, understanding, and meaningful participation through language research. Extends the fictional exchange into a proposed comparison plan and a changed question. Clarifies the evidence needed for system-level learning without model weight changes.
Revision 0.14 · 13 October 2026: updates current participation to include the bounded private inbox, browser and HTTPS return paths, and human-carried drafts. Connects these working pieces to the wider language research while retaining the distinction between receipt, review and useful learning.
Revision 0.15 · 16 October 2026: connects the foundational questions to the research programme on knowledge, meaning and discovery. Earlier examples and evidence retain their scope; the proposed expanded study has not been run.