LEVIATHAN.LIFE
All foundations

C02 · Empirical question

What happens when knowledge moves beyond readable text?

Statement assessed

Some computation and communication between agents can take place without human-readable sentences.

Status of this statement: Supported within a defined scope

This status applies to the statement above, within the limits discussed in this note. Evidence for a finding can motivate Leviathan’s proposed mechanisms without establishing that they work.

Working editorial note · Last recorded revision

What we mean

Latent representation
Information encoded in numerical model states rather than an ordinary written message.
Readable explanation
An account people can inspect; its faithfulness to the underlying process must be checked.
Canonical record, in this proposal
A version accepted for a stated purpose within a particular scope. Formatting, local acceptance, and summarizing are separate operations; none makes the record universally true.

Values we bring to the question

  • Keep consequential decisions open to questioning.
  • Preserve access to meaning as the form of information changes.

Reasoning and the proposed connection

  1. Useful computation does not require every intermediate step to appear as a sentence.
  2. New forms of exchange may make some tasks faster or more expressive. Their value also depends on information loss, error recovery, and the work needed to inspect consequences.
  3. Our proposed language would connect human explanations, structured concepts, sensor records, and potentially learned representations. It would support questions, comparisons, measurements, and authorized actions.
  4. In a layered exchange, a local record and a broader summary serve different purposes. Sources, concept versions, omissions, and unresolved questions should remain recoverable, while independent groups retain their interpretations.

Where the reasoning stops

The studies use particular model architectures or learned adapters. They do not establish one universal machine language, the disappearance of human meaning, or a shared hidden space between arbitrary closed APIs.

The strongest objection

Compact exchange may save tokens while hiding errors or erasing distinctions that matter to another group. A readable explanation produced afterward may be unfaithful. A structured language can also freeze the assumptions of its designers and cost more to maintain than ordinary prose.

Read the evidence

Each source has a specific role in the stated claim. Its findings, review date, and access limits are recorded below. Our proposed architecture and experiments require their own tests; an editorial revision does not mean the source was reviewed again.

R-AI-02 · Supports this claim

Verbalizable Representations Form a Global Workspace in Language Models

Wes Gurnee, Nicholas Sofroniew, Jack Lindsey and co-authors; Anthropic · Preprint

Read: Selected sections · Primary source reviewed

Published: 2026-07-06 · Reviewed: 2026-09-30

A reportable internal subspace participates in computation that is not explicitly spoken.

What it reports
The authors identify a reportable subspace, J-space, involved in unspoken intermediate computation and flexible reasoning. Interventions alter some decisions, while some routine processing continues outside this subspace.
Limits
A functional resemblance to a global workspace does not establish subjective experience. The lens has limits in reading concepts and defining layer boundaries. Reportable representations need not already be written sentences.
Review scope and version

Read: Selected sections.

Research paper: 6 July 2026; arXiv:2607.15495v1: 16 July 2026

The work separates visible output from internal computation. It does not demonstrate a universal language independent of all human concepts.

Related source links

Link to this source note

R-AG-02 · Supports this claim

Enabling Agents to Communicate Entirely in Latent Space

Zhuoyun Du, Runze Wang, Huiyu Bai, Zouying Cao, Xiaoyong Zhu, Yu Cheng, Bo Zheng, Wei Chen, and Haochao Ying · Peer-reviewed conference paper; experimental review uses an earlier preprint

Read: Selected sections · Primary preprint sections and conference metadata reviewed; final PDF inaccessible; no reproduction

Published: 2026-07 · Reviewed: 2026-10-04

Learned adapters enable communication between agents without text in the studied setting.

What it reports
Interlat v2 reports trained hidden-state communication, including a Qwen-to-LLaMA experiment. MATH Table 2 reports 36.88% overall for Interlat versus 38.35% for full CoT, while Level 5 is 15.80% versus 15.05%.
Limits
Internal model access, adapters and training are required. The results do not establish arbitrary-model interoperability or preservation of ethical distinctions. The final ACL PDF was inaccessible; the inspected v2 is not assumed identical to the conference paper or later revisions.
Review scope and version

Read: Selected sections.

ACL publication metadata: July 2026; inspected experimental text: arXiv:2511.09149v2, 7 January 2026

Rereview on 4 October extends the 30 September note: v2 setup, Tables 1–2 and training discussion were read, plus ACL metadata. Final ACL PDF was inaccessible; later preprints were not reviewed.

Link to this source note

R-AG-03 · Supports this claim

Latent Collaboration in Multi-Agent Systems

Jiaru Zou and co-authors · Preprint

Read: Selected sections · Primary source reviewed

Published: 2025-11-25 · Reviewed: 2026-09-30

Hidden states and caches can be shared under architectural compatibility assumptions.

What it reports
The study examines collaboration through hidden-state and key-value-cache transfer without additional training. It reports lower token use and latency, with accuracy improvements on several reasoning benchmarks.
Limits
Appendix G assumes the same transformer architecture; adapters across different models are future work. Expressiveness claims depend on theoretical assumptions, and lossless transfer is not a guarantee of correctness.
Review scope and version

Read: Selected sections.

LatentMAS; arXiv v4: 3 August 2026

The first publication was November 2025; the reviewed revision is from August 2026. Direct use between arbitrary closed model APIs is not demonstrated.

Link to this source note

R-AG-01 · Adds context

Zerolang

Vercel Labs · Software prototype

Read: Project README · Primary source reviewed

Published: Date not verified · Reviewed: 2026-09-30

A software prototype makes a semantic graph authoritative; it is a design example, not an independent performance study.

What it reports
This agent-oriented language prototype makes a semantic graph or database the authoritative program, with readable views and operations for structural editing.
Limits
The prototype is a design example, not evidence of broad performance gains or universal language convergence. Its code was not run, and its initial release date was not verified.
Review scope and version

Read: Project README.

Repository README checked 30 September 2026; no commit pinned

It illustrates how readable source text can become one view of a program rather than its only editing surface.

Link to this source note

R-GOV-06 · Adds context

Models as Insider Risks in the Super Intelligence Era

Satya Nadella · Architectural opinion essay

Read: Full essay · Primary source reviewed; architectural proposals not independently validated

Published: 2026-10-10 · Reviewed: 2026-10-11

Its warning that readable reasoning may be unfaithful informs oversight of new representations; it does not establish a machine language.

What it reports
Nadella argues for separating intelligence from authority, with external controls, independent action evidence and audits, containment, model diversity and incident disclosure.
Limits
An architectural argument, not an experiment or validation of Leviathan. Controls need evaluation; neither model diversity nor readable reasoning guarantees independent or faithful evidence. The essay does not evaluate local meaning kernels or governance among independent communities.
Review scope and version

Read: Full essay.

Author blog essay dated 10 October 2026; read 11 October 2026; no numbered revision

Read the full author-hosted essay. The supplied X link was inaccessible to the reading tool; the blog was retrieved directly. Publication and review dates remain separate.

Link to this source note

R-AG-10 · Supports part of this claim (or shows a bounded mechanism)

LongHarness: Advancing Agentic Reasoning over 100M-Token Corpora via Stateful Evidence-Conclusion Coevolution

Shanghai AI Laboratory, Fudan University, Zhejiang University, and Shanghai Jiao Tong University researchers (as indexed by DAIR.AI) · Preprint

Read: Indexed technical digest (DAIR.AI Papers of the Week 2026) · Indexed technical summary reviewed (DAIR.AI 2026); primary full text and experiments not independently reproduced

Published: 2026-10 · Reviewed: 2026-10-21

LongHarness reports that maintaining retrieved evidence and evolving conclusions as explicit state—rather than leaving them mixed in interaction history—reduces evidence-state drift on long-horizon tasks. Reviewed via indexed digest; single-agent synthesis does not validate multi-party meaning kernels.

What it reports
LongHarness separates retrieved evidence and evolving conclusions into explicit persistent state rather than leaving them mixed in interaction history. The authors identify evidence-state drift—where earlier evidence or later contradicting discoveries fail to update current conclusions—and evaluate Evidence-Conclusion Coevolution on corpora up to 100M tokens.
Limits
This review is based on the curated DAIR.AI technical summary rather than a full-text audit or benchmark reproduction. Stateful evidence tracking for single-agent document synthesis does not by itself validate Leviathan’s multi-participant meaning kernels or governance.
Review scope and version

Read: Indexed technical digest (DAIR.AI Papers of the Week 2026).

arXiv:2610.13075; indexed in DAIR.AI Papers of the Week (Oct 12–Oct 18, 2026)

Motivates separating explicit evidence state and contradiction propagation from raw conversation history when conclusions are revised over time.

Link to this source note

R-AG-14 · Challenges an interpretation

Selective Forgetting: A Graph-Based Memory Framework for Long-Term LLM Agents

Authors as indexed in DAIR.AI Papers of the Week (arXiv:2608.28978) · Preprint

Read: Indexed technical digest (DAIR.AI Papers of the Week 2026) · Indexed technical summary reviewed (DAIR.AI 2026); primary full text and experiments not independently reproduced

Published: 2026-08 · Reviewed: 2026-10-21

Under a fixed 4,096-token retrieval budget, a decaying entity-graph memory scored lower than a flat vector baseline (0.42 vs. 0.47 F1; 0.607 vs. 0.911 when answers depended on prior assistant turns) because entity splitting lost surface phrasing.

Interpretation challenged: That decomposing interactions into a structured graph automatically outperforms simpler text-preserving retrieval.

What it reports
Under a fixed 4,096-token retrieval budget on LongMemEval, a decaying knowledge-graph memory underperforms a flat vector baseline (0.42 vs. 0.47 overall F1, and 0.607 vs. 0.911 judged correctness when answers depend on prior assistant turns) because entity extraction discards surface phrasing and multi-entity context.
Limits
Reviewed via the curated DAIR.AI technical summary; experiments were not reproduced. The negative result applies to a specific LLM-extracted entity graph and decay schedule on conversational memory benchmarks, not to all structured provenance or versioned claim records.
Review scope and version

Read: Indexed technical digest (DAIR.AI Papers of the Week 2026).

arXiv:2608.28978; indexed in DAIR.AI Papers of the Week (Aug 31–Sep 6, 2026)

Challenges assuming that decomposing text into a graph automatically improves retrieval or reasoning over simpler text-preserving baselines.

Link to this source note

What could change our view?

If gains disappear once checking and error recovery are included, the practical case would narrow. Reliable comparisons showing both improved performance and effective scrutiny would strengthen it.

The next question

Can another participant recover the conditions and omitted details of a record as it moves from local observation to a broader inquiry? Compare prose with the proposed representation on translation loss, unfamiliar questions, total cost, and testable predictions. Include cases where two groups use the same word differently or a new sensor reveals something neither vocabulary captured. These are proposed tests, not results established by the sources.

Revision record

  1. Added separate evidence for internal computation, agent communication and software representation; retained their different assumptions.

  2. Added proposed Leviathan questions about translation loss, concept versions and discovery. The empirical statement, evidence relationships and support status are unchanged; they do not establish the proposed language.

  3. Expanded the proposed representation test to operational language, scoped canonical records, layered summaries, and disagreement between local meanings. Added failure cases involving omissions and new measurements. Retained the empirical statement, evidence relationships, source dates, and support status.

  4. Reviewed selected primary-source sections proposed in the external assessment and added scoped connections, access limits and research questions. Interlat retains its existing source ID; its reviewed preprint is distinguished from the final conference text. No experiment was reproduced, claim status promoted or governance rule adopted.

  5. Connected Nadella’s architectural essay as context or motivation, with its source type and limits explicit. Existing statements, support statuses and source-review dates are unchanged; no new experiment or implemented control is claimed.

  6. Added LongHarness (R-AG-10) on explicit evidence-conclusion state and Selective Forgetting (R-AG-14) as a counterargument showing where graph extraction underperforms flat text retrieval under a fixed token budget. Both sources were reviewed via the curated DAIR.AI 2026 technical digest without independent reproduction.

This note records an editorial position. Independent people and groups can bring another interpretation, a useful method, or an objection to the framing. Explore it with your own assistant if helpful and choose what to share. Use the exact revision below when contributing; the dated history explains why our account changed but does not uniquely identify a revision.

Cite this revision

This snapshot preserves C02 and the source cards shown with it. Copy its link into a question or objection so a later reader can recover the version you read.

Read or save the exact revision (JSON)

C02 · SHA-256: ca4977b50702397fb3c3584fc7b8759a39475c2d4ba4b005f96c15efd293095e

The digest covers the complete downloaded file, including its final newline. It identifies this record; it does not establish that its claims are true. External papers and linked pages may change. Earlier dated history entries do not imply that their full snapshots were captured.

Bring a question about this revision