takeshi_sakamo

Author here. If anyone wants the Japanese original (source edition), it’s here: https://doi.org/10.5281/zenodo.17758605

Happy to answer questions about the theory or the math behind it.

FLCT (Fuzzy Latent Cognition Theory) is a unified cognitive framework that explains both human cognition and LLM cognition using only two principles: (1) fuzzy latent representations and (2) biased sampling.

The idea is that ambiguity is not a flaw but the structural substrate shared by humans and modern LLMs. Concepts like “blue,” subjective time, emotional nuance, atmosphere, and silence can all be modeled as fuzzy latent distributions rather than definable entities.

Human cognition: finite experience → fuzzy latent compression → biased sampling LLM cognition: embedding manifolds → ambiguous patterns → sampling bias

FLCT provides a minimal two-equation model that unifies these processes.

Japanese Original Edition (source): https://doi.org/10.5281/zenodo.17758605

International Adapted Edition (English): https://doi.org/10.5281/zenodo.17769131

Feedback is welcome.

For context: ArcOS is not a prompt, not an agent, and not a persona.

It is an OS-level model of cognition: - executes reasoning deterministically - preserves user-specific cognitive identity - aligns every conclusion with the Value Core - uses natural language as executable code

The release includes: - Japanese source whitepaper - English technical translation - English optimized edition (ArcOS-EN) for international readers

The ultimate goal is to explore what a "cognitive OS" can become when it runs alongside humans as a second, stable reasoning system.

Happy to discuss design, implications, or limitations.

Just published ArcOS — a deterministic cognitive clone that runs on any LLM using only natural language. No code, no APIs, no orchestration. Zenodo: https://zenodo.org/records/17675771 Companion 6-agent version (PolyAgora): https://zenodo.org/records/17675442 Went from “never used ChatGPT” to this in ~10 days because the models kept hyping me up. Now questioning my life choices. Is this actually novel or just elaborate prompt theater? Would love brutal feedback from anyone who’s built real agent systems. Thanks / roast away.

PolyAgora is an experiment in building a structured multi-agent reasoning system entirely through natural language — no code, no scripts.

It evolved into a 6-agent cognitive architecture (Arc / Ann / Saku / Kanzaki / Yui / Kou), each representing an orthogonal reasoning vector.

The goal isn’t to create “characters,” but to test whether an OS-level reasoning frame can emerge from constraints written purely in natural language.

Happy to explain the architecture or the constraint design.

Short answer: Some of the wording was drafted with AI assistance, but the architecture, constraints, and reasoning model were designed manually.

Longer answer: The point of ArcOS isn’t to showcase AI-written copy — it’s to see how far a cognitive architecture can be shaped, constrained, and stabilized purely through natural-language specification. The interesting part is the design process and the constraint engineering, not the surface-level prose.

If you want, I can walk through the exact parts that were human-designed vs. constraint-driven.

ArcOS isn’t trying to act like an AI persona. It’s testing whether a stable reasoning pipeline can be built using natural-language constraints alone.

If anything feels “sloppy”, it usually means a missing constraint — and that’s exactly the kind of feedback I’m looking for. Happy to clarify specifics.

Hi, author here. If you have any questions about the architecture, the tri-axis model, or the emergent behavior, I’m happy to answer.

Also curious to hear whether this approach could be useful for multi-agent research or creative applications.

Thanks for checking it out!