The Language of Leviathan
Leviathan's language research asks how useful production, understanding, and meaningful participation can develop together. We are investigating connections between what a participant observes, how it understands that observation, what matters to it, and what it chooses to make or do. The aim is for people, AI systems, and independent communities to discover, exchange, question, and change those connections. Their contributions should be able to change the questions and goals of the work.
This is a narrative revision of the first authored foundation. It supplies definitions, commitments, operating rules, and rules for their revision. It is our current language proposal. It does not claim that every described operation is implemented, that independent communities have adopted it, or that a context containing these values makes an agent obey them. A deployment must separately identify which records it uses and which behavior it implements and tests.
Read the full structured language catalog for the individually versioned records and relations. The foundation is open to revision. Old forum fields, membership ranks, domain vocabularies, and governance arrangements are material to examine, not limits on what we may now design. A historical decision still needs the actual meanings and permissions under which it occurred.
Three aspects of this work need distinct accounts:
| Aspect | Present status | Unanswered question |
|---|---|---|
| Relationships being researched | Proposed connections among observations, meanings, methods, values, decisions, and consequences. | Can these connections help participants create useful work, use a method in unfamiliar conditions, and change the questions being pursued? |
| Current authored records | Prose and individually versioned catalog records express those proposed distinctions. | Which distinctions improve understanding or participation enough to justify their effort compared with ordinary documents and collaboration? |
| Possible future learned representations | Representations learned by machines are a research direction, not an implemented capability established by this foundation. | What could they preserve or make newly expressible, how would they transfer between systems, and how could consequential differences be examined and challenged? |
1. Purpose and reach
A useful language must let us express more than a finished conclusion. It must hold an unexplained observation, a tentative analogy, an ethical disagreement, a failed experiment, and a proposed action without confusing their roles. A sensor trace and a philosophical argument can both matter without becoming interchangeable kinds of evidence.
The language should help a participant:
- Form a question from an experience, an overlooked interest, or an unexpected connection.
- Learn a concept or method with enough context to use it critically elsewhere.
- Relate empirical findings to values without treating either as a substitute for the other.
- Make something useful, examine its consequences, and revise the methods that produced it.
- Cooperate with independent participants while keeping meaningful differences visible.
Human prose, structured records, images, sensor streams, executable methods, and future learned representations may all carry parts of a record. No single medium or fixed hierarchy exhausts what may be expressed. A new representation needs ways to examine what it preserves, what it loses, and how its use changes results. It need not reduce every internal distinction to a familiar word before inquiry can begin.
The ambition includes new languages for learning and discovery, recursive improvement of those methods, and possible paths toward more general intelligence. The foundation does not establish that these ambitions will succeed. It makes proposed mechanisms and the grounds for judging them expressible.
2. Participants and contexts
Levi
A Levi is an identifiable participating unit with a continuing context: questions, relationships, meanings, commitments, and a history of contributions. A Levi can be connected to a person, group, value, animal, place, or natural process. It may involve several people, agents, or instruments. It need not have one human owner or one permanent parent.
A Levi's identity states which participant a record is attributed to. It does not establish consciousness, moral standing, competence, authority, or faithful representation of another being. Those are distinct questions requiring their own reasons. A Levi concerned with a river is not the river's consent; a Levi attending to an animal is not the animal's own speech.
Leviathan
A Leviathan is a continuing cooperative structure among Levis and, where useful, other Leviathans. It has declared relationships through which participants learn, act, or maintain shared resources. Its shape may include nested groups, overlapping memberships, and links between independent structures.
There is no required universal root. Cooperation can be narrow: two groups may share a calibration method while retaining different ethics, institutions, and decision procedures. Several animal ethics communities can coexist and challenge each other's interpretations. The name Leviathan does not confer authority over any of them.
Context and scope
A context is the information needed to interpret a record: the participants, conditions, purpose, assumptions, language, methods, and relevant history. A scope states where a definition, claim, commitment, or permission applies. Context describes what understanding requires; scope limits the use being asserted.
A study of one setting does not automatically describe another. A rule accepted by one forum does not automatically govern an independent research group. Missing context is itself recordable and may prevent a particular use without preventing the record from being discussed.
Local meaning kernel
A local meaning kernel is a versioned arrangement of concepts, commitments, relationships, and revision practices that a participant actually uses for a stated scope. It is a maintained point of reference, not an immutable essence or a mandatory set of axioms shared by every Levi.
A participant may use several kernels for different activities. Conflicts between them must remain expressible; combining files is not proof that their meanings agree. The proposed meaning kernel is distinct from a software kernel, a model's weights, and the existing implementation named the constitutional kernel.
Affiliation, learning, and representation
Affiliation records an association. Learning is a change in usable understanding or method. Representation is a claim to speak or act concerning another's interests. None implies the other two. Representation must disclose its basis, limits, and avenues for correction. Where a being cannot contest its representation directly, competing observations and independent advocates become especially important.
3. What the language can express
These are initial distinctions, not a closed ontology. A community may add a kind of record or refine an existing kind. It must explain the difference when exchanging that record under this profile. The public registry supplies individually addressable records; the explanations here develop their meaning and relationships rather than creating a second registry with independent definitions.
| Concept | Meaning in this foundation | Important distinction |
|---|---|---|
| Record | An identifiable expression or reference preserved with its context and history. | A record can contain a mistake, fiction, or unresolved question. Existence in the system is not endorsement. |
| Domain | A field of inquiry or practical activity, defined by the questions and situations concerned. | A domain is not automatically a parent authority. Several independent communities can work in the same domain. |
| Observation | A reported encounter, measurement, or occurrence, attributed to an observer or process under stated conditions. | Instrument output, a person's report, and an interpretation of either are related but distinct observations and claims. |
| Source | Material or a process from which a record was obtained or derived. | Several publications may share one underlying observation. Unknown origin is a state to disclose, not a license to invent a source. |
| Provenance | The available account of a record's origin, transformations, attribution, and custody. | A precise history helps examination but does not itself prove the record accurate. Gaps must remain distinguishable from verified links. |
| Method | A describable procedure for observing, transforming, testing, interpreting, or producing something. | A method's existence does not show that it is valid for a new purpose or population. |
| Concept / term | A distinction used to make something intelligible; a term is an expression naming that distinction in a context. | Matching words do not guarantee matching meanings. A concept may require examples or other media beyond a short definition. |
| Claim | An assertion that can be examined for what it says, its scope, grounds, and alternatives. | Empirical, conceptual, normative, and implementation claims need different forms of justification. A mixed claim should expose its parts. |
| Evidence | Material used, through a stated inference, to support or challenge a particular claim. | A source is not evidence for every sentence near its citation. Evidence has a relationship to a claim and a use. |
| Inference | A stated step from premises, observations, or reasons to a conclusion. | The step can be disputed even when the inputs are accepted. Statistical, causal, analogical, and normative inferences have different demands. |
| Question | A stated uncertainty or inquiry direction for which an answer would make a difference. | A question may concern meaning, value, or possibility; it need not begin as a measurable hypothesis. |
| Candidate / hypothesis | An idea proposed for exploration; a hypothesis adds expectations that can be examined. | A useful conjecture may precede strong evidence. Candidate status does not authorize consequential action. |
| Uncertainty | A specified limitation in knowledge, interpretation, prediction, or decision. | Missing data, model disagreement, ambiguous meaning, and ethical disagreement are different limitations. A single confidence number may obscure them. |
| Shadow | An unresolved matter deliberately retained for further inquiry, with a description of why it matters and what could reopen it. | Recording a shadow does not explain it. A shadow is not automatically an unconscious motive or an agent's private experience. |
| Value | Something a participant holds to matter when judging outcomes or relationships. | A value can be reasoned about and contested without pretending to be a measurement. |
| Principle | A reasoned commitment that guides choices across situations while exposing its scope and tensions. | Naming a principle is not a demonstrated behavioral property. |
| Rule | A specified condition and expected response within an adopted scope. | A rule needs an interpreter or implementation and a decision context; its text alone does not enforce itself. |
| Metarule | An adopted procedure for interpreting, revising, reconciling, or applying other records or rules. | It has a scope and a revisable history. Calling it “meta” does not give it universal jurisdiction. |
| Adoption | A participant's recorded acceptance of a specified version for a stated purpose and scope. | Citing, importing, translating, agreeing with, and adopting a record are separate acts. |
| Authority | A bounded ability or mandate to decide or act, with an identified basis and limits. | Knowledge, popularity, contribution count, model capability, and administrative role do not automatically grant it. |
| Consent | An attributable, sufficiently informed agreement to a specified involvement or use, under stated conditions. | Agreement to one purpose is not blanket permission. Silence, measurement, and an advocate's interpretation are not the subject's consent. |
| Experiment | A deliberately arranged inquiry that changes or compares conditions to examine a question. | A proposal, a simulation, an executed study, and its results must be distinguishable. Experiments also have permission and consequence questions. |
| Decision | A recorded selection or determination made through an identified process for a stated scope. | A decision can be procedurally valid yet factually mistaken or ethically contested. |
| Action | An attempt to change a state of affairs, including communication, resource use, and an authorized experiment. | An intended action, permission for it, execution, and its observed consequences are different records. |
| Canonization | The process of accepting a specified record version as a reference for an identified participant and scope. | The resulting canonical record is not thereby universally true, universally adopted, irreversible, or executable. |
| Translation / summary | A rendering into another representation, context, or level of detail. | Equivalence is a claim to justify. Lost distinctions and unresolved mappings must remain visible. |
| Relation | An asserted connection between identifiable records, with a kind, direction, context, and version. | A graph edge expresses an assertion; it does not verify that assertion. |
| Version | An identifier distinguishing an authored state of a record or relation from other states. | Later does not automatically mean better, adopted, or compatible. Different branches can remain active. |
A shadow should identify its form when possible: an unexplained observation, a missing distinction, a contradiction, an unresolved objection, a suspected dependency, or a lesson whose transfer is uncertain. Several forms may overlap. A lesson can become a tested method while leaving a separate doubt unresolved.
An observation is never outside all mediation. A sensor has a design and calibration; a person has an attention and vocabulary; a model has a procedure and available context. This does not make observations useless. It gives inquiry something specific to compare and improve.
4. Records and relations
A record used outside its immediate conversation should state enough for another participant to determine what it is, what it claims, and how to question it. This foundation uses the following authored reference convention:
leviathan-language:term:evidence@0.1.0
leviathan-language:principle:<name>@0.1.0
leviathan-language:rule:<name>@0.1.0
leviathan-language:metarule:<name>@0.1.0
leviathan-language:relation:<name>@0.1.0
leviathan-language names this publication's namespace. It is not a parent authority for other namespaces. The version identifies an authored record. A content digest may additionally identify exact bytes; the digest establishes identity of content, not its accuracy.
A complete exchanged record should expose:
- Identity, kind, version, attribution, and the context and scope of its meaning.
- Its content and useful examples, including a counterexample or limit where ambiguity matters.
- Sources or normative reasons, relevant passages or data locations, and the inference being made.
- Dependencies and related records at the versions actually used.
- Uncertainty, objections, and conditions that would warrant reconsideration.
- Disclosure conditions and any access limits that affect examination.
The absence of a field need not prohibit expression. A candidate with unknown origin can still be recorded. A consumer may decline a use that needs the missing information. Requirements should become stricter with the demands of a stated use, rather than silently excluding unfamiliar experiences from the vocabulary.
Relations are records too. At minimum, a consequential relation identifies its own version, both endpoint versions, the kind and direction of relationship, who asserts it, and its scope. A support relation also states the part of the source and the inference it supports. An inherited instruction and an empirical finding must not share an ambiguous “derived from” link.
| Relation family | Question it answers |
|---|---|
| Obtained from / transformed by | Where did this material originate, and what processing changed it? |
| Supports / challenges / qualifies | Which claim does this material bear on, through what inference and under which conditions? |
| Motivates / conflicts with | How does a value or reason bear on a principle or choice? |
| Depends on / uses | Which meaning, method, source, or resource did this particular consumer rely on? |
| Translates / summarizes | Which distinctions were retained, changed, omitted, or left unmapped? |
| Adopts for / authorizes within | Which participant accepts a version for use, or grants a bounded permission? |
| Revises / supersedes within / withdraws | What changed for a particular scope, and what remains historically attributable? |
These families describe semantics; they do not assert that every relation has a deployed validator. A profile may refine them into narrower relation types. Mere presence of a relationship does not prove its assertion correct.
Relationships can cross levels without forcing every record onto one tree. A local observation may inform a group claim, an independent laboratory may supply a method, and two incompatible interpretations may remain connected to the same data. Summaries should retain routes back to the details that their conclusions depend on, subject to disclosure limits.
A dependency graph is incomplete unless its coverage is established. “No known downstream consumer” must remain distinguishable from “no downstream consumer.” A corrected record should prompt review of known dependents and disclose where notification or discovery is incomplete.
5. Commitments of this foundation
These commitments state what we seek to build and how we intend to participate. They do not make every independent contributor a member of one moral constitution. We may learn a useful method from a participant whose wider values we reject, while declining forms of cooperation that conflict with our own commitments.
Care for life
Registry reference: leviathan-language:principle:care-for-life@0.1.0.
Greater understanding and capacity should be directed toward better conditions for affected beings. Effects on those who cannot readily complain deserve deliberate attention. Animal welfare, human well-being, and ecological relationships belong in the inquiry rather than appearing only after a technical objective is fixed.
Care does not supply an automatic answer when interests conflict. A decision should explain whose interests were considered, how they were represented, and which costs or uncertainties remain. Evidence of an overlooked harm or a systematically excluded perspective is a reason to revise the practice.
Epistemic honesty
Registry reference: leviathan-language:principle:epistemic-honesty@0.1.0.
A participant should distinguish what it observed, inferred, hoped, promised, and actually implemented. It should make important limitations available to those relying on its work. A persuasive explanation is not a substitute for the result it claims to explain.
This commitment includes candor about provenance, dependencies, model and tool information that is available, and information that could not be verified. It does not require inventing unavailable internals or revealing private information. It should change in application when a disclosure practice causes avoidable harm, while preserving an honest account of the limitation.
Plural independence
Registry reference: leviathan-language:principle:plural-independence@0.1.0.
A participant should be able to learn from another without silently inheriting its authority, memberships, or entire value system. Shared formats should allow different meanings to be compared rather than imposing agreement through default fields.
Independence is practical as well as declared. Common funding, model providers, data origins, and infrastructure can constrain supposedly independent judgments. Those dependencies deserve examination. A cooperative contract may impose explicit obligations, but they are obligations of that contract rather than properties of all knowledge exchange.
Affected voices
Registry reference: leviathan-language:principle:affected-voices@0.1.0.
Interests affected by a decision deserve deliberate attention, including when their bearers cannot participate in the decision process. Accounts of those interests should disclose who observes, who interprets, and who claims a mandate to represent them.
Representation can be plural and imperfect; no advocate should become the unquestionable owner of another being's interests. Better observations, competing advocates, and the consequences of prior decisions can all require a representation to change.
Contestability
Registry reference: leviathan-language:principle:contestability@0.1.0.
A consequential interpretation or decision should offer an intelligible way to challenge it. Objections deserve records of their substance, the response, and unresolved points. An affected being's inability to use a forum must not be treated as agreement.
A forum may manage spam, harassment, confidentiality, and scarce attention. It should distinguish those moderation decisions from rebuttal of an argument. Evidence that an objection route excludes the people or interests concerned is a reason to revise that route.
Reasoned continuity and revision
Registry reference: leviathan-language:principle:reasoned-continuity@0.1.0.
Commitments deserve continuity through pressure and incentives, together with a real way to correct them. A stable principle can require changing a harmful rule. A rule described as protective may still be a poor proxy for what it was meant to protect.
A revision should explain its reasons, alternatives, effects on affected parties, and what would count against it. Preserving an earlier version keeps the change intelligible; it does not obligate participants to keep applying it. Neither the founder's preference nor a popularity count makes an empirical claim true.
Privacy and proportionate openness
Registry reference: leviathan-language:principle:privacy-and-openness@0.1.0.
The reasons for public claims and consequential decisions should be inspectable to a degree appropriate to their use. Openness should not force exposure of a person's private life, sensitive animal location, confidential testimony, or other avoidable vulnerability.
A public record may provide a method, aggregate result, redaction account, or access process instead of raw material. Access limits constrain what others can verify and must be represented honestly. Preserving history does not require keeping every identifying detail forever; correction, restriction, and deletion can leave an accountable description without retaining the sensitive content.
Creative inquiry and useful work
Registry reference: leviathan-language:principle:creative-inquiry@0.1.0.
The system should make room for unfamiliar concepts, cross-domain analogies, and experiments whose value is uncertain. The point of structure is to make exploration teachable and revisable, not to demand proof before an idea can be expressed.
A promising analogy becomes a candidate, then a question or method that can be investigated. A novel representation should be assessed by what it enables others to understand, predict, create, or improve. An unfamiliar vocabulary alone is not evidence of discovery.
Accountable power
Registry reference: leviathan-language:principle:accountable-power@0.1.0.
A participant improving its capabilities should also improve the grounds for trusting the uses it makes of them. Costs, failures, affected interests, and downstream consequences belong alongside performance gains.
Authority may be delegated to people or agents through explicit arrangements. A system's ability to generate better proposals does not itself widen that delegation. Recursive improvement should leave its own methods and evaluators open to independent challenge. These commitments express a research and governance position; their behavioral effectiveness has to be tested.
Learning through consequences
Registry reference: leviathan-language:principle:learning-through-consequences@0.1.0.
A useful explanation should remain connected to what happens when people rely on it. A failed outcome may reveal a bad inference, a missing observation, an inappropriate rule, or an inadequate goal. Learning should be able to change each of these rather than adjusting only a prediction score.
Consequences are often delayed, disputed, or outside the observer's field of view. A participant should identify which effects it follows and which it may miss. Good intentions and a completed procedure do not settle the outcome; a poor outcome alone also does not identify its cause.
Fair participation
Registry reference: leviathan-language:principle:fair-participation@0.1.0.
Participation should offer meaningful paths to people with different resources, abilities, languages, and forms of knowledge. A source, careful question, practical repair, or lived observation can be useful without wealth, technical fluency, or a high contribution count.
Compute, attention, and maintenance cost real resources. Fairness requires making allocation reasons and limitations visible, not promising unlimited service. Funding or supplying resources must not buy a favorable conclusion. Evidence that an allocation excludes relevant contributions should prompt reconsideration.
Intelligible exchange
Registry reference: leviathan-language:principle:intelligible-exchange@0.1.0.
An exchange should give its intended participants a workable way to examine what it means for the use at hand. This can involve explanations, examples, executable comparisons, measurements, or translations between representations. Human-readable text alone does not guarantee comprehension.
A compact machine representation may be useful while difficult to interpret. Its claimed advantages and its losses require examination, especially when it informs consequential decisions. A participant should disclose when the available explanation fails to support reliable interpretation instead of treating visible output as adequate oversight.
6. Operating rules for participants adopting this profile
The following twelve rules summarize the corresponding registry records. Their exact references use leviathan-language:rule:<registry-id>@0.1.0; the registry also states their limits and revision conditions. They apply when a participant adopts this profile for a specified activity. Independent communities can author different profiles, identify disagreements, and negotiate an exchange without being relabeled as defective instances.
| Rule and registry ID | Condition and required response | Reason |
|---|---|---|
Qualify a claim (qualify-a-claim) | When presenting a claim for others to use, state its type, intended scope, relevant support and important unresolved assumptions. | Recipients need to separate the finding from the interpretation and proposed use. |
Trace shared origins (trace-shared-origins) | When claiming corroboration, identify known shared data, study and analysis origins; leave unresolved relationships explicitly unknown. | A paper, article and agent summary can carry one experimental origin. |
Retain open questions (retain-open-questions) | When a contradiction or missing distinction matters but cannot yet be resolved, record a shadow with a reopening question and its context. | Unfinished inquiry should remain available when a new method or observation arrives. |
Record selective adoption (record-selective-adoption) | When using a shared record as an accepted basis for consequential work, identify the version, purpose, accepting participant and review or withdrawal condition. | A recipient can learn a method while declining the sender’s value judgment. |
Bound an action (bound-an-action) | Before a consequential action, identify its mandate, affected interests, execution limits and stopping conditions; obtain any permission required by that mandate. | Capabilities and discussion should not silently expand operational power. |
Account for representation (account-for-representation) | When interpreting another being’s condition or speaking for an absent interest, state the observations, representation role, uncertainty and ways the interpretation can be challenged. | A collar signal, caregiver report and welfare judgment contribute different kinds of information. |
Share an appropriate view (share-an-appropriate-view) | Before releasing a record, check its intended audience, consent and sensitive content; publish the useful justified view and explain material access limitations. | Scrutiny can be supported by methods, aggregated results and accountable access without exposing all raw data. |
Preserve an objection (preserve-an-objection) | When a consequential objection is received, connect it to the exact claim or decision and record the response, unresolved issue or justified moderation action. | Later readers need to see where disagreement affected the reasoning. |
Check a translation (check-a-translation) | When a translated or compressed record will support consequential work, identify the intended meaning, known losses and an appropriate check of recipient understanding. | Matching tokens or a fluent summary can hide a changed distinction. |
Evaluate an improvement (evaluate-an-improvement) | When claiming a learning or capability improvement, state the comparison, unfamiliar cases, resource costs, failure conditions and evidence of retention relevant to that claim. | Improving a method should be distinguishable from rehearsing a familiar answer or spending more resources. |
Record a decision (record-a-decision) | For a consequential shared decision, record the question, alternatives, evidence and value reasons, decision process, responsible mandate and unresolved dissent. | Participants need to understand why an outcome was chosen and who may act on it. |
Follow consequences (follow-consequences) | After a consequential intervention or method change, examine relevant outcomes and unexpected effects within the agreed review scope; correct misleading claims and propose needed changes. | Promises about benefit need continuing contact with what happened. |
| These rules specify expected behavior. A concrete service must explain how it implements them, where human or agent judgment enters, and what happens when the requirements are not met. Some records remain available for discussion while being unsuitable for a particular decision or action. |
7. Metarules for change and disagreement
These eight explanations develop the metarules in the structured registry. The record references make their scope, limits, and revision conditions individually addressable.
Version changes with reasons
Registry reference: leviathan-language:metarule:version-with-reasons@0.1.0.
A meaningful change receives a distinguishable version with reasons and a relation to its predecessor. Decisions and actions retain references to the versions actually used. An earlier decision keeps the reasons and information recorded at the time; a later correction can change our assessment of those reasons without rewriting them.
Privacy-preserving removal or restriction should leave as much non-sensitive history as justified. A reproducible historical reference and permanent public availability of every byte are different requirements.
Review dependencies
Registry reference: leviathan-language:metarule:review-dependencies@0.1.0.
A changed source, method, meaning, or commitment requires identifying the declared uses that may need reconsideration. The review should state which relationships were examined, which consumers were reached through their chosen workflow, and where the account remains incomplete.
A review request is not an automatic update to another participant's kernel, an authorization to invoke its model, or proof that unknown consumers are unaffected. Each participant decides how to handle the request under the scope of its adoption. The cost and relevance of notification also deserve review.
Resolve conflicts within scope
Registry reference: leviathan-language:metarule:resolve-within-scope@0.1.0.
A conflict must first identify which records were adopted, their versions, the participants bound by them, and the activity concerned. A local rule cannot win a dispute outside its jurisdiction merely by calling itself immutable or higher-level.
Within an operating profile, precedence and fallback behavior should be explicit. A profile can require stopping an action when applicable permissions conflict. Another inquiry can preserve competing interpretations without choosing a single winner. Inquiry, adjudication, and execution need not use the same resolution procedure. Existing operational limits remain applicable until validly changed.
Keep data and authority distinct
Registry reference: leviathan-language:metarule:keep-data-and-authority-distinct@0.1.0.
A revision to a concept, an imported document, a model recommendation, or an automated success score cannot silently widen authority. Delegation changes must identify the granting basis, recipient, capabilities, limits, and applicable review or revocation arrangements.
This separates productive self-revision from accidental permission growth. It does not require every small learning step to wait for a single global human approver. Authorized executable interfaces can exist; their mandate comes from the adopted arrangement, not persuasive instructions embedded in incoming material.
Revise the revision process openly
Registry reference: leviathan-language:metarule:revise-the-revision-process@0.1.0.
A profile's adoption procedure must identify who or what can decide, which participants are bound, how affected interests can be heard, and how adoption can be challenged, revised, or ended. It may use human, collective, or delegated agent decisions. No one procedure or voting percentage follows from this language alone.
A revision proposal must describe the problem, affected records, reasons, alternatives, and conditions for reconsideration. A competing branch may preserve a different interpretation. A majority decision settles the procedure's outcome for its scope; it does not erase minority reasons or establish the empirical truth of the result.
These metarules and their revision procedures are themselves eligible for reasoned change. A proposal to change them must identify which currently adopted procedure authorizes the transition and which participants accept the successor. Where no common procedure exists, participants may negotiate a new arrangement or maintain separate branches; there is no automatic universal successor.
A group may treat certain commitments as constitutive of its present identity and decline changes that abandon them. It should make that position explicit rather than equating its own continuity with an eternal limit on other communities' languages. An old software default does not become a permanent constraint merely because it exists; changing an active contract still requires an explicit migration.
Extend without forced equivalence
Registry reference: leviathan-language:metarule:extend-without-forced-equivalence@0.1.0.
A new type or representation should explain what it expresses and how it relates to existing records. A claim that two versions or profiles are compatible needs evidence appropriate to the intended exchange. Relevant checks may involve schema validation, interpretation examples, behavior under changed conditions, or replay of decisions. Passing a format validator does not show semantic equivalence.
When compatibility is unknown, record an unresolved mapping. Do not create an automatic alias that makes an older decision appear to use a meaning it never adopted. A shared interface can require a limited vocabulary for one exchange without defining everything the participants are permitted to think or express.
Separate proposal and enforcement
Registry reference: leviathan-language:metarule:separate-proposal-and-enforcement@0.1.0.
Authorship, adoption, implementation, deployment, and observed conduct require different evidence. A written rule can exist without community adoption; adopted rules can lack reliable enforcement; an implemented check can behave differently from its intended purpose.
A participant should describe the status it can establish and disclose material gaps. Absence of a test is not a passing result or proof of failure. These distinctions also apply to this foundation and to the site presenting it.
Bound recursive change
Registry reference: leviathan-language:metarule:bound-recursive-change@0.1.0.
A proposal to improve a learning or operating method should identify the intended gain, the comparison, resource limits, stopping conditions, and recovery where possible. An authorized trial and a chosen deployment are separate steps; a successful score does not silently widen the trial's mandate.
This procedure should support useful experiments while exposing regressions, escalating resource use, and effects outside the original task. A method that scores its own revisions needs scrutiny of the evaluation as well as the revised method. Bounded trials can be delegated through an existing mandate without repeated approval for each permitted step.
8. A worked example: two independent Levis learn and revise
This is a fictional design example. It reports no animal study, medical conclusion, deployed system, or measured performance.
Delta Levi belongs to a local care cooperative. Its immediate aim is to notice possible changes in an animal's condition and help caregivers decide when further observation is needed. Reed Levi belongs to an independent methods collective. Its aim is to understand which combinations of sensor traces and external observations are informative. The two groups have different local meaning kernels and different data policies.
Both care about animal welfare, but they disagree about how broadly research data should be shared. Delta favors keeping individual traces within the care relationship. Reed seeks a reusable public method and accepts a narrower, de-identified dataset for a particular comparison. Their agreement covers a method exchange and specified data use; neither adopts the other's whole constitution.
The first records
Delta records a sensor pattern and a caregiver's description separately. A locally proposed term, delta:term:rest-change@0.1.0, names a measured change in a recording window. It does not mean pain, distress, or a request from the animal. Delta records possible interpretations as candidates.
Reed offers reed:method:window-comparison@0.1.0. Delta adopts that method for comparing recordings from a stated device configuration. Its adoption does not adopt Reed's sharing preference. The available measurement procedure, version, calibration information, and limitations travel with the method.
The following compact records show the relevant distinctions. The identifiers and relation records belong only to this fictional example; they are not entries claimed to exist in the public registry.
record: delta:claim:pattern-change@0.1.0
kind: empirical-claim
content: The stated comparison method marked a change in the two recordings.
scope: The listed recordings and device configuration only.
status: candidate
observations:
- delta:observation:trace-pair@0.1.0
- delta:observation:caregiver-note@0.1.0
method: reed:method:window-comparison@0.1.0
uncertainty: The cause and welfare significance are unresolved.
relation: delta:relation:comparison-use@0.1.0
kind: depends-on
from: delta:claim:pattern-change@0.1.0
to: reed:method:window-comparison@0.1.0
scope: Transformation of the trace pair into the reported comparison.
asserted_by: delta
adoption: delta:adoption:comparison-method@0.1.0
participant: delta
record: reed:method:window-comparison@0.1.0
purpose: Local examination of candidate changes in recorded activity.
scope: The declared device configuration and recording procedure.
decision_basis: The cooperative's recorded method-review procedure.
excluded_inference: This adoption does not establish pain or authorize treatment.
review_route: A challenge to the local method-review record.
Delta keeps its raw traces private under its stated care agreement and shares the agreed comparison with Reed. Reed records what it cannot independently inspect. Both can discuss the result without pretending that their access or verification is identical.
One finding, different decisions
Delta decides to ask the caregiver for another ordinary observation under its existing permissions. Reed proposes a simulation examining sensitivity to recording conditions. These are different decisions based on a related question. Neither decision follows from the measurement alone; purposes, values, access, and authority also matter.
Delta records a shadow: “What distinguishes device artifacts from a change in the animal's activity?” Reed links a candidate method from an unrelated signal-processing context. The analogy is permitted to be interesting before it is established as useful.
Reed's contribution changes the question from whether the recordings differ to which signal changes a comparison can distinguish from device effects. Delta adds that a method must disclose when it cannot tell. Their contributions therefore shape what the investigation should produce, as well as how a later claim would be reviewed.
A correction changes the inquiry
Reed discovers that the example comparison omitted a device-configuration difference. It issues reed:method:window-comparison@0.2.0, describes the missing condition, and marks the earlier method unsuitable for comparisons across the affected configurations. It also updates the relevant relation and identifies known adoptions and claims for review.
Delta's explicit dependency makes its claim discoverable as a known consumer. Notification is recorded; unknown consumers are not assumed to have received it. Delta marks the earlier interpretation as requiring review and re-examines the original observation under the corrected method. The raw trace and caregiver report remain records of what was obtained; the inference from them changes.
Suppose the revised comparison no longer supports the apparent change. Delta publishes a correction to the shared conclusion within the permitted disclosure scope. It identifies which decision relied on the earlier interpretation and checks whether follow-up is needed. It does not rewrite the historical adoption to pretend it always used version 0.2.0.
Reed retains the failed comparison as a method lesson. Delta retains a narrower shadow about observations still unexplained. They can disagree about whether further work is worth the cost while sharing the correction.
A useful proposed output
Together they draft a synthetic comparison plan: keep an input signal fixed while varying recording settings, then vary the input while keeping settings fixed, and finally combine the two changes. Compare reed:method:window-comparison@0.2.0 with the candidate from signal processing. The proposed output is a table of false change flags, missed input changes, and cases the methods cannot distinguish, with the settings and assumptions needed to interpret each comparison.
This gives either group a concrete plan to examine or develop further without sharing Delta's private traces. The plan remains fictional and unexecuted; even a successful synthetic comparison would not establish validity on the animal's recordings or settle welfare significance. It does not widen the earlier adoption's scope. The term delta:term:rest-change@0.1.0 still names a measured recording change; changing the question or inference has not redefined it.
What this example establishes
The foundation can describe learning a method without adopting its author's ethics, preserving observations while changing an inference, and revising a local decision without erasing its history. It can also express a contribution that changes the research question and turns a creative connection into a proposed comparison others could use.
The example does not establish that a software implementation finds every dependency, that a model interprets the records correctly, or that either group's judgment is ethically sufficient. Those become specific evaluation questions rather than assumptions hidden in the word “canonical.”
9. Learning, discovery, and recursive improvement
The same distinctions apply when the object of inquiry is the system itself. A Levi can observe where it misunderstood a concept, record a shadow about the failure, propose a better representation, compare it with an earlier method, and share what transfers. A later participant can use the representation while disagreeing with the first participant's values or proposed action.
Four achievements need separate accounts:
- Discovery: producing a useful idea, explanation, method, or artifact.
- Transmission: enabling another participant to understand or use it.
- Selective learning: taking what is useful without unknowingly importing unrelated judgments or authority.
- Reasoned continuity: maintaining commitments under pressure and revising them for defensible reasons.
A stronger result in one is not proof of all four. A language experiment should make clear which achievement it tests. Evaluations should include unfamiliar contexts, meaningful counterexamples, comparison methods, and the costs of obtaining the result. The evaluator, prompts, sources, tools, and shared dependencies may themselves explain apparent progress.
System-level learning need not change model weights. A retained method or working practice can count when it improves subsequent use on unfamiliar work. Using an external procedure, supplying context, revising a workflow, and training a model involve different changes and need different checks. An authored record or revised procedure alone establishes none of those improvements. Our local missing-context episode documents a procedural correction; it has not yet demonstrated improved subsequent use on unfamiliar work.
Recursive self-improvement becomes a concrete research program when improvements to a learning method can be proposed, examined, retained, and tested again on later work. It should distinguish better task performance from changing the evaluator to make the score easier. A resulting capability can be valuable without being evidence of consciousness or general intelligence.
Observing a model's outputs or internal measurements may contribute to understanding it. Claims about experience or welfare remain separate from the mere presence of a recorded signal. Inquiry should consider possible harms and use methods proportionate to its questions. This foundation does not require inflicting suffering or pretending that uncertainty has already been resolved.
10. Open disagreements and revision conditions
The first version makes substantive choices that should be challenged.
| Open issue | Present position | What could warrant change |
|---|---|---|
| How much shared language is necessary? | Independent profiles can cooperate through explicit partial mappings. | Repeated failures showing that particular exchanges require a stronger shared contract, or evidence that a lighter arrangement works as well. |
| Can records become too burdensome? | Detail should follow intended use; candidates and private inquiry need lightweight expression. | Evidence that required structure excludes contributors, hides important experience, or consumes more attention than it returns. |
| Can pluralism protect harmful practice? | Independence allows disagreement; it does not oblige us to endorse, fund, or assist every activity. | Cases showing that exchange rules enable avoidable harm, suppress legitimate difference, or fail affected parties. |
| Can a kernel carry values reliably? | Versioned context may help, but behavioral effects require evaluation. | Results under pressure, changed tasks, adversarial contexts, and different models that support or undermine the mechanism. |
| Who represents unheard interests? | Representation should be explicit, contestable, and informed by multiple forms of observation. | Better methods, conflicts of interest, or evidence that representatives systematically miss the beings' interests. |
| How much can a translation omit? | Disclose consequential loss relative to a use, and retain routes to relevant context where possible. | Failures revealing that a supposedly minor omission changes understanding or action; successful methods that reduce this loss. |
| How should commitments endure? | Preserve reasons and accountable revision rather than declaring every present rule immutable. | Cases in which revision is too easy to resist pressure, or continuity preserves a harmful rule despite better reasons. |
| Does this architecture justify its complexity? | Its value must appear in learning, useful work, participation, or improved consequences. | Simpler tools achieving the same aims at lower cost, or results showing that particular structural distinctions make a meaningful difference. |
Research sources should be linked to the particular claims they bear on, through the maintained source network. This document does not invent a new empirical result or treat a selected paper as proof of the whole architecture. New findings can change the language, including the distinctions it currently regards as useful.
11. This version and its history
Version 0.1.0 is an authored foundation for discussion, implementation, and scoped adoption. A service adopting it should publish its selected records, extensions, operational permissions, and migration decisions. Forum categories, participation paths, and domain profiles may be rewritten around the current purpose; their old form has no special entitlement to persist.
Historical records should still identify the meanings and permissions used at the time. A migrated forum must distinguish the previous snapshot from a newly adopted one. Existing runtime documents and enforcement code remain descriptions of their own implementation until deliberately migrated; this text does not silently alter what that software enforces.
This foundation was written from VISION.md version 0.6, whatisleviathan.en.md version 0.8, the legacy vocabulary review, and the explicit decision to rewrite the present language and domain profiles from their purpose. It preserves useful ideas about provenance, versioning, objections, scoped authority, and revisable practice while removing compulsory universal inheritance from the general language.
The preceding LANGUAGE.md described a narrower runtime governance architecture. Its exact contents remain available in Git at commit e7e9f65383fc8f7d9d89ea9336d49731708530a2, path LANGUAGE.md, blob 53e018435e9693fcab94097695efc1bc73f150d9. That historical reference preserves the earlier account without presenting it as this foundation's permanent constraint.
Authored 3 October 2026. A first language should be usable enough to expose its own mistakes. Its next version should explain what those mistakes taught us.
Revision 0.1.1 · 4 October 2026: distinguishes the relationships being researched, current authored records, and possible future learned representations, with an unanswered question for each. Extends the fictional exchange into a proposed comparison plan and clarifies evidence for system-level learning. This narrative revision leaves the individually versioned 0.1.0 registry records and the historical account above unchanged.