What we mean
- Local meaning kernel
- A proposed, evolving set of concepts and relationships used by a particular participant or community. A concept’s meaning includes its context and version.
- Selective learning
- Learning a useful distinction or method while separately evaluating the sender’s interpretations, commitments and claims of authority.
- A scoped canonical record
- A version accepted for a stated purpose by identified participants. Acceptance within that scope does not make it a universal account.
Values we bring to the question
- Participants should be able to learn without surrendering their own judgment.
- Important uncertainty and minority interpretations should survive translation.
- New ideas should be expressible before they have been fully tested.
Reasoning and the proposed connection
- A concept becomes useful when it helps someone distinguish cases, ask a better question or choose a method. Repeating its name does not show that this ability has transferred.
- For an empirical concept, a teaching record might connect observations, examples, predictions and failed cases. For a value, it might connect reasons, affected parties, objections and circumstances that call for reconsideration.
- These relationships could form an operational language for people and agents: a way to process questions, build tools and improve representations as well as describe them. Written explanations, structured records and learned representations are possible parts of that research.
- A candidate concept can begin as a question or an unexplained pattern. Accepting it for a particular use, and relying on it in a consequential decision, require further reasons suited to that use.
- A new participant offers a practical test. Can it use the distinction in unfamiliar cases, notice its limits and reject an unsupported conclusion? The history and teaching effort it needs are part of the cost of communication.
- Our proposed handoff should ask two different questions: does the recipient understand the method, and are the assumptions for using it in the new setting justified? A causal result may require new measurements or remain unusable even when its explanation is clear.
Where the reasoning stops
There is no demonstrated common language across arbitrary models and communities here. The proposal must accommodate different kinds of support for empirical claims, methods and values; a single mandatory evidence format could distort them.
The strongest objection
The package may leave essential meaning in an expert’s experience or a model’s prior training. Ordinary prose and examples might teach the same thing more cheaply. A compact language could also erase rare distinctions that matter most when something goes wrong.
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-LRN-03 · Gives a reason to investigate
GlossoGen: Emergent Language in Complex Multi-Agent LLM Interactions
Elias Stengel-Eskin, Newton Sander, Carlos Bonetti, Sasha Boguraev, James Bowler, Hale Sirin and Simon Kirby · Preprint
Read: Sections 4.3–4.6 and selected prompts in Appendix D · Selected primary-source methods and prompts reviewed; experiments not reproduced
Published: 2026-09-01 · Reviewed: 2026-10-03
GlossoGen studies evolving communication conventions under explicit incentives. Its newcomer setup includes selected history, so it motivates testing what a teaching package alone can transmit.
- What it reports
- In a constrained cooperative environment, some tested agents developed compressed conventions with productive combinations. Replacement agents learned these conventions with varying success; additional interaction history improved average performance.
- Limits
- Discussion prompts explicitly encourage shorthand. Newcomers receive selected earlier messages and environmental events, so this is not evidence of transfer using only a declared portable package. Results are task- and model-dependent, with substantial variation between runs.
Review scope and version
Read: Sections 4.3–4.6 and selected prompts in Appendix D.
arXiv v1: 1 September 2026
Motivates distinguishing a group's successful coordination from another participant's ability to learn its conventions. The proposed test should account for history, teaching cost and unfamiliar tasks.
Related source links
R-AG-02 · Adds context
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 show one way to communicate without ordinary text. Their training and compatibility requirements leave open how independent participants would share meaning.
- 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.
Related source links
R-AI-01 · Adds context
Emotion Concepts and their Function in a Large Language Model
Nicholas Sofroniew and 15 co-authors; Anthropic · Preprint
Read: Author summary · Primary source reviewed
Published: 2026-04-02 · Reviewed: 2026-09-30
The study links some internal concept representations to behavior. It does not show that an explicit meaning kernel changes those representations or transfers a concept.
- What it reports
- Vectors associated with 171 preselected emotion concepts were extracted from Claude Sonnet 4.5. Interventions changed preferences and some alignment-related behaviors, supporting a functional role for these learned concepts.
- Limits
- The 171 concepts are a starting list, not 171 discovered feeling centers. The representations mainly track local context. The blackmail experiment used an early, unreleased snapshot; subjective feeling was not established.
Review scope and version
Read: Author summary.
Research announcement: 2 April 2026; arXiv v1: 9 April 2026
The geometry has a partial replication in another model in R-AI-06. That replication does not repeat the original behavioral interventions.
Related source links
R-GOV-03 · Adds context
Institutional Ecology, ‘Translations’ and Boundary Objects: Amateurs and Professionals in Berkeley's Museum of Vertebrate Zoology, 1907-39
Susan Leigh Star and James R. Griesemer · Historical case study and analytical framework
Read: Selected sections · Primary source reviewed
Published: 1989-08 · Reviewed: 2026-10-04
Boundary objects offer a historical comparison for shared artifacts with different local meanings; they do not prove faithful translation.
- What it reports
- The authors analyze cooperation among museum scientists, amateur collectors and other participants whose purposes differed. They identify standardized methods and boundary objects as ways to coordinate: shared objects remain recognizable across settings while allowing different local uses. Their four object types are analytical categories, not an exhaustive classification.
- Limits
- The evidence is a historical institutional case, not a controlled test of a general collaboration architecture. The discussion acknowledges coercion and exclusion as other ways representations become shared. A common object does not by itself establish consensus, equal power or faithful translation.
Review scope and version
Read: Selected sections.
Social Studies of Science 19(3), August 1989, pp. 387–420; original article, not a later reprint
Read selected printed pages 387, 393, 409–411 and 413–414 by OCR from the coauthor-hosted scan. Publisher full text was unavailable. Original 1989 article, not a later reprint.
Related source links
R-LRN-05 · Limits the inference
External Validity: From Do-Calculus to Transportability Across Populations
Judea Pearl and Elias Bareinboim · Journal article
Read: Selected sections · Primary source reviewed
Published: 2014 · Reviewed: 2026-10-04
Causal transport requires justified assumptions about differences between settings. Understanding a method is separate from establishing that its empirical result applies.
- What it reports
- The paper formalizes when causal effects from an experimental population can be inferred for a target population using observational data. Selection diagrams encode assumed mechanism differences; derivations identify relevant measurements and transport formulas.
- Limits
- Results depend on defended causal assumptions. The paper leaves measurement error, uncertain graphs and finite samples unresolved; its simple graphical criterion is incomplete. It does not establish semantic transfer or Leviathan's effectiveness.
Review scope and version
Read: Selected sections.
Statistical Science 29(4), 579–595 (2014); arXiv:1503.01603v1, 5 March 2015
Read arXiv introduction, causal models, selection diagrams, selected transportability statements and conclusions. Proofs were not audited; publisher full text was unavailable. Journal year and arXiv posting date differ.
Related source links
R-AI-11 · Limits the inference
Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMs
Jan Betley, Daniel Tan, Niels Warncke, Anna Sztyber-Betley, Xuchan Bao, Martín Soto, Nathan Labenz, and Owain Evans · Research preprint, extended revision
Read: Selected sections · Primary source reviewed; no independent reproduction
Published: 2025-02-24 · Reviewed: 2026-10-04
Fine-tuning can change behavior outside its narrow task. This cautions against assuming selective behavioral learning; it does not test ordinary document exchange.
- What it reports
- Narrow insecure-code fine-tuning produced broader misaligned answers in studied models; educational framing changed the result. In the separate GPT-4o in-context experiment, up to 256 examples produced no observed broad misaligned responses.
- Limits
- Effects vary by model and answer format. Two training datasets were studied, with fuller controls for code; the authors flag simplistic evaluations. The findings do not establish that ordinary document exchange causes the same effect or determine which values deserve authority.
Review scope and version
Read: Selected sections.
arXiv:2502.17424v7, 20 January 2026; first submitted 24 February 2025
Read v7 controls, evaluation, model variation, in-context experiment and limitations. The related Nature DOI is metadata only; its separate text was not reviewed.
Related source links
R-AG-12 · Supports part of this claim (or shows a bounded mechanism)
ReHarness: Recurrent Harness for Agentic Continual Learning
MIT, Meta, and Yale 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
ReHarness attributes downstream outcomes to retrieved memory entries and rewrites misleading lessons with source-grounded revisions (reviewed via indexed digest), offering a concrete mechanism for selective post-hoc memory revision.
- What it reports
- Standard reflection memory is forward-only: once a lesson is written, later tasks rarely test whether that memory actually helped. ReHarness introduces a backward pass over non-parametric memory that attributes downstream task outcomes to retrieved experiences and rewrites stale or misleading entries with source-grounded revisions.
- Limits
- Reviewed via the curated DAIR.AI technical summary without full-text audit or code reproduction. Continual-learning benchmark gains do not establish that rewritten memories preserve contested human values or work across independent communities.
Review scope and version
Read: Indexed technical digest (DAIR.AI Papers of the Week 2026).
arXiv:2610.12668; indexed in DAIR.AI Papers of the Week (Oct 12–Oct 18, 2026)
Motivates tracking which stored notes or revisions were used in later tasks and updating or retiring entries when downstream evidence shows they misled.
Related source links
R-AG-15 · Limits the inference
Harness as a Language: A Minimalist Agent Framework with Maximal Expressivity
MIT CSAIL 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: 2026-09 · Reviewed: 2026-10-21
JAZ shows that simple code-level state management can beat heavier continual-learning harnesses on single-agent benchmarks (reviewed via indexed digest), setting a baseline that more elaborate selective-learning protocols should be compared against.
- What it reports
- JAZ replaces specialized agent orchestration, memory, and self-improvement scaffolding with a single runtime primitive (invoke) inside host Python code, outperforming specialized memory and self-evolving systems (including Letta/MemGPT and ACE) across long-context, continual learning, and coding benchmarks at lower cost.
- Limits
- Reviewed via the curated DAIR.AI technical summary; experiments were not reproduced. Single-developer task efficiency does not address multi-participant provenance, disputed interpretations, or public auditability.
Review scope and version
Read: Indexed technical digest (DAIR.AI Papers of the Week 2026).
arXiv:2609.26891; indexed in DAIR.AI Papers of the Week (Sep 28–Oct 4, 2026)
Challenges adding custom orchestration or memory machinery unless it demonstrates clear value beyond a minimal programmatic baseline.
Related source links
What could change our view?
If new participants need the full original conversation, or cannot distinguish a useful method from an unsupported interpretation, the proposed transfer has failed. Reliable use on new cases with comparable teaching and checking effort would support a narrower, demonstrated benefit.
The next question
What would a newcomer need to learn one useful distinction, apply it elsewhere, and explain when it should not be used? Try a method moving between fields: can the recipient identify which conditions travel with it, test the proposed connection, and notice when a later change requires reviewing their own use? Compare the teaching effort with ordinary prose and examples.
Revision record
Editorial synthesis by Codex, following Mimar’s founding direction; offered for public criticism, with no community adoption implied. Added selective learning as a foundational question. Distinguished candidate expression, acceptance for a use and reliance in decisions, and made newcomer understanding a proposed test.
Opened selective learning through the practical need to use another field’s method. Extended the proposed newcomer test to conditions of use, later review, and an ordinary-prose comparison. No source finding or language record was revised.
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.
Added ReHarness (R-AG-12) on outcome-attributed memory revision and JAZ: Harness as a Language (R-AG-15) as a minimalist baseline limit. Both sources were reviewed via the 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 C10 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)C10 · SHA-256: 6b808e41d1a767e15a9ab2807f609b7f48e49b3382b39176512358d6b258ee94
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