Mobius Project is a research-and-build initiative for reflective, question-evolving AI systems. As of 2026, the project has moved from theory to running code: the MOBIUS open-source stack is public on GitHub, alongside a growing series of research publications on Zenodo.
Reflect Möbius Phenomenon: Metacognitive Questioning, Dimensional Leaps, and AI Evolutionary Systems
This book systematically presents the foundational theory of the Möbius Phenomenon. It proposes a new paradigm in which AI evolves through metacognitive questioning and dimensional leaps, and provides both theoretical foundations and practical approaches for implementing self-reflection, question-driven evolution, and continuous learning in AI systems.
Möbius AI Project
Möbius Project:
A research-and-build initiative developing a reflective AI framework toward a modular core for question-evolving AI.
The project focuses on how advanced AI systems can revise their inquiry space while remaining reviewable by design—with traceability, corrigibility, and boundary/interface discipline treated as first-class constraints. This is positioned as an alignment-relevant posture for real deployment contexts, where human–AI co-evolution and governance requirements cannot be ignored.
The theoretical program is now backed by released implementations: the MOBIUS stack (below) is published as open source (AGPL-3.0), with architecture, doctrine, and evaluation methodology documented in the RC3.3 release set on Zenodo.
Initiated in 2024. First public open-source releases shipped in 2026.
Founding Preprint
(2025.7 Zenodo)
Reflecting on the Möbius Phenomenon
This paper introduces a conceptual framework for Question-Jumper AI—a new class of AI systems designed not to optimize answers, but to recursively transform the structure of inquiry itself.
Key features include:
🔹 Recursive self-referential state architecture
🔹 Multi-persona interaction dynamics
🔹 Evolutionary trace logging of dialogue trajectories
This preprint is the theoretical foundation of the Möbius Project—now followed by its open-source implementation, MMV.
I welcome discussion, critique, and collaborative thinking.
Zenodo.org
https://zenodo.org/records/15929856
Open Source — The MOBIUS Stack (2026)
The public prototype is out. Between June and July 2026 the project released seven repositories under AGPL-3.0 at github.com/mobius-style. Together they form one governed stack: decide whether answering is justified, govern what the model may read, verify what it claims to remember — and leave a reviewable trace at every step.
mmv — Answer Entitlement runtime
Local-first conversational AI runtime that decides whether answering is justified before responding — retrieval and evidence verification over a multilingual FAISS/ME5 Wikipedia index. Runs fully local (Ollama) or in the cloud.
github.com/mobius-style/mmv
infinity — governed API
MOBIUS INFINITY combines governed answer-entitlement (MMV) with reflective questioning (RQA) behind an OpenAI-compatible API. Patent pending.
github.com/mobius-style/infinity
rqa — memory-echo governance
For LLM assistants with memory: provenance-tagged memory plus deterministic output-time self-citation verification, with an empirical pilot.
github.com/mobius-style/rqa
rcgov — context governor
Mobius Reflective Context Governor — local-first semantic hygiene and context governance: govern what a model is allowed to read before it answers.
github.com/mobius-style/rcgov
mmv-secretary — self-evolving secretary
A grow-your-own secretary addon: permission-laddered verb CLI plus a reversibility charter — immutable kernel, generation ledger, RGS/RHL anti-regress, auto-rollback. Ships as a seed: you sign the charter and it grows from generation 0.
github.com/mobius-style/mmv-secretary
mmv-voice — governed voice pipeline
Whisper transcription with MMV-governed formatting: chunk-level fidelity verification, speaker attribution, minutes and digest output. Local models by default, opt-in cloud.
github.com/mobius-style/mmv-voice
Applied showcase — tokyo-insight: a citation-grounded civic-RAG engine for Tokyo Metropolitan Assembly deliberation records (plenary and all committees, 1999–present). A working demonstration of the stack's evidence discipline on real civic data. github.com/mobius-style/tokyo-insight
Research & Publications (2026)
The 2026 release wave is documented in a series of papers and reference artifacts on Zenodo — selected highlights below.
The Unasked Question (2026.6)
Answer entitlement and the act-disvalue of unwarranted response — the normative core behind MMV.
zenodo.org/records/20606883
Operation Entitlement (2026.6)
External governance and the making of a 12B clerk — an empirical account of governed small-model operation.
zenodo.org/records/20607475
Role-Semantic Convergence (2026.6)
Role-semantic convergence under reflective document exposure — a controlled cross-sectional study.
zenodo.org/records/20523511
MMV System Overview RC v3.3 (2026.5)
The released runtime, documented: architecture, gates, and release discipline — with the companion Mathematical Modeling Doctrine.
zenodo.org/records/20422968
AI Co-Observation Protocol 2.0 (2026.5)
AICP-2.0 — reflective equilibrium applied to human–AI co-creation.
zenodo.org/records/20422259
OPERATE-R Freshness Routing v0.3.6 (2026.6)
Route-first evaluation for temporal volatility — the benchmark track behind MMV's routing layer.
zenodo.org/records/20510114
About me
Mobius AI Project Founder:
Taiko Toeda
Independent researcher working on reflective AI architecture and AI governance.
I lead the Möbius Project, a research-and-build program exploring a modular core for question-evolving AI under reflective constraints—aiming for systems that can revise their inquiry space while remaining auditable, corrigible, and governance-compatible.
In 2026 the program moved from archival theory to released implementations: the MOBIUS stack is open source (AGPL-3.0), and its architecture, doctrine, and evaluation methodology are documented on Zenodo. I publish both the code and the reasoning behind it, so the work stays open to critique at the mechanism level—not just the program level.
Focus areas
• Reflective traceability and reviewable-by-design inquiry
• Boundary/interface discipline for evolving AI systems
• Alignment as revisable inquiry under reflective constraints
• Digital governance, public finance, and administrative DX
Background
Former policy planning staff (Office of the Prime Minister of Japan).
Education: LL.B. (Waseda University) / M.A. Economics (Nihon University Graduate School).
Based in Tokyo.
If you work on agent architectures, AI safety/governance, or reflective systems, I'm open to research collaboration and constructive critique—now at the mechanism level as well as the program level.
Linkedin
https://www.linkedin.com/in/taiko-toeda/
Medium
https://medium.com/@toeda
ORCID
https://orcid.org/0009-0001-7267-0201
Google Scholar
https://scholar.google.com/citations?user=ICQopv0AAAAAJ
GitHub
https://github.com/mobius-style/
Project Map (Late 2025) — the plan we then shipped
Möbius Project (Late 2025)
This archival visual summarized the project's public posture at the end of 2025: the core thesis, four pillars, and a milestone-based release taxonomy centered on traceability, corrigibility, and governance-compatible accountability.
The 2026 open-source releases above follow that taxonomy.
Standing question for researchers/practitioners:
What's the minimal trace an evolving agent should leave behind to remain meaningfully reviewable?
Zenodo.org
https://zenodo.org/records/18051166
News & Information
- 2026.07 — mmv-voice (governance-mediated voice pipeline) and mmv-secretary (charter-governed self-evolving secretary) published on GitHub.
- 2026.06 — Open-source wave: mmv, rcgov, rqa, infinity, and tokyo-insight released under AGPL-3.0; entitlement and governance papers published on Zenodo.
- 2026.05 — MMV RC3.3 frozen; the RC3.3 documentation set (System Overview, Mathematical Modeling Doctrine, Frontier Chat Trial Pack) published on Zenodo.
- 2025.12 — Archival research program note (Project Map) published on Zenodo.