Mobius AI Project

Flagship Paper — The Unasked Question (2026.6)

The Unasked Question: Answer Entitlement and the Act-Disvalue of Unwarranted Response. The normative core of the Möbius program: before asking whether an answer is correct, ask whether answering is justified at all. This is the thesis the MMV runtime implements in code, the flagship books carry into everyday language, and the 2026 measurement campaign tests empirically.
Zenodo.org — https://zenodo.org/records/20606883

Flagship Paper (Verification) — Termination by Novelty (2026.8)

Termination by Novelty: A Machine-Checked Theory of Anti-Regress Guards for Reflective Governance Processes. The Unasked Question states what the program claims; this states how a claim is shown to hold. For a self-revising process under a novelty guard — every accepted transition must be semantically non-zero, and no-change must be declared explicitly — the theorems say the process cannot loop forever while appearing to progress: silent stagnation becomes a terminating, detectable state. Proposition 1 and both termination theorems are machine-checked in Lean 4 (core, no external libraries; the tightness of the Theorem 1 bound is an elementary hand proof). The checked artifact ships with the paper, so the proof can be re-run rather than taken on trust.
Zenodo.org — https://zenodo.org/records/21837211

New — An evaluation plan for a composite financial stress indicator, specified but not run (2026.10)

An Evaluation Plan and Design Study for a Composite Financial Stress Indicator (v1.0.0, deposited 18 October 2026) is a software-and-text record: an evaluation plan specified in advance and not run, the code to run it, and a design and power study on synthetic series — which constraints of the specification act, what moves its adaptive weights, and what a plan with five test episodes could detect. Chapter 8 of the revised 2026 edition of Reflective Economics & Financial Engineering cites it as Toeda (2026k).

The record is exploratory and contains no empirical result and no market data. The plan was not run, by the author's decision; the reasons are stated in its README and in Chapter 8. Nothing in it is evidence that the indicator discriminates stress from calm, warns of stress, or is better than, worse than, or as good as any baseline. Code is AGPL-3.0-or-later, text CC BY-NC-SA 4.0. The study was made with AI tools under the author's direction and checked by separate AI runs instructed to refute it, and by models from two other developers; no person independent of the author has examined it, and no independent party has reproduced it. zenodo.org/records/23448619

Testing Revisable Authority: the empirical companion (2026.10)

Testing Revisable Authority (v1.0, deposited 3 October 2026) is the preregistered empirical companion to From Warranted Acts to Revisable Authority. Its protocol, code, seeds and thresholds were hashed and timestamped before any confirmatory run. The technical component implements the five procedural specifications D1–D5 as a standalone reference layer and met every pre-specified criterion on the five confirmatory seeds, under scripted trajectories with planted ground truth. The institutional component traces complaint handling under Articles 20 and 21 of the EU Digital Services Act for Meta, TikTok and Google (YouTube) from public records, with two coders of different model families.

The institutional component is exploratory, by the study's own frozen rule: inter-coder agreement in the pilot was κ = 0.092, below the 0.60 gate, and no further pilot provider existed in the frozen record. On a clarified codebook the main coding reached κ = 0.773, reported as an exploratory estimate. The decision, evidence and funding links were mostly visible in the dispute-settlement bodies' public reporting; the duty and remedy links mostly could not be verified from public records. These are statements about what public records show at these dates, not findings of compliance or non-compliance. No human participants; no human referee. The record includes the code, all traces and both coders' sheets. zenodo.org/records/23110635

From Warranted Acts to Revisable Authority (2026.10)

From Warranted Acts to Revisable Authority: Practical Judgment, Critique, and the Continuing Justification of AI Delegation (v1.0, deposited 3 October 2026) is the third paper in the answer-entitlement series. The Unasked Question asked when an AI system is warranted in answering; the second paper asked what infrastructure supports bounded, recoverable delegation; this one asks whether a sequence of individually warranted and recoverable acts can justify the continuing arrangement that selects their purposes and the criteria by which they are judged, and answers that it cannot. Where a consequential arrangement defended as supporting a person's agency helps shape the endorsements, preferences or evidence later cited for its continuing authority, those items cannot by themselves certify that authority unless their admissibility remains open to the person's challenge by an effective route whose outcome does not depend solely on the criteria in dispute. The paper withdraws three sufficiency readings of the program's own earlier position and states five procedural specifications, D1–D5, as design hypotheses.

It is a critical synthesis and internal correction of the program, not a new general theory, and it reports no empirical measurement of its own. Its section 10.5 summarizes the confirmatory technical results of a separately preregistered companion study, in which D1–D5 were implemented as a reference layer and met every pre-specified criterion under scripted trajectories with planted ground truth; that shows implementability as specified and nothing about real users, deterrence, agency over time or enforceability. The companion study is in preparation and will carry the protocol, code and data. The text was reviewed only by AI agents; no human referee has reviewed it. zenodo.org/records/23103483

Semantic Interface Model and three further records (2026.10)

Four records were deposited on Zenodo on 1–2 October 2026: three preprints and one companion data-and-code record. Each states its own limits, and they are repeated here.

• Semantic Interface Model v1.0 — the classical core under MUSE 2.0 and TCR 3.0. A channel is modeled as a Markov chain from a task variable to an internal representation, and its semantic energy is the Shannon mutual information I(Y; Z). A scalar and a tensorial semantic aperture are built on it, and cognitive curvature is defined as the curvature of a task-weighted Fisher information metric, with worked examples checked by a script. No theorem in the paper is new, and no measurement on a deployed system is reported. zenodo.org/records/23064318
• Half-Step Negotiation v1.1 — a formal account of the half-step principle: a reflective AI system should not habitually complete the next conceptual move on behalf of the user. The Half-Step Target Estimator treats the choice of a next question as a constrained optimization over a finite candidate set. The paper does not fix the feature functions or the thresholds; its quantities are proxies, and it is refutable through hypotheses it states. zenodo.org/records/23086312
• A sharp escape-rate bound on information flow between two binary variables — a note in stochastic thermodynamics. For a stationary Markov jump process in which two binary variables never jump at the same time and every allowed transition has an allowed reverse, the information flow between them is bounded in terms of a cap on the escape rate of every state and a cap on the entropy production rate, and an explicit four-state chain attains the bound. A self-contained script checks the constants and the attainment. The literature search was narrative and non-systematic, the note makes no priority claim, and no human referee has reviewed the text. zenodo.org/records/23081365
• Reflective Light Cone V1.3 Local Validation Supplement — companion data and code for MUSE and Bounded Reflection: a deterministic simulation and six arms of a 100-item operational track, with a README of reproduction steps and ten reading notes. Supplemental evidence only: not an independent benchmark and not a proof of the Reflective Light Cone conjecture. zenodo.org/records/23072189

Unanimity in Disguise (2026.9)

Two detectors running in the Mobius stack — the saturation signal of the Reflective Budget Governor and the saturation score of the Reflective Homeostasis Layer — score a short window of observations and fire at a threshold of 0.70. This note proves that at every window length up to 18 that condition is exactly unanimity: the threshold is not measuring degrees of agreement at all. The result is a sharp envelope argument, not a benchmark. zenodo.org/records/22842506

What a Retrieval Index Loses Before Anyone Notices (2026.9)

A forensic re-read of three Wikipedia embedding indexes the project had published itself and was running in production — English, Japanese and Chinese, FAISS over multilingual-e5-large. Three defects, none of which had ever produced an error message.

• What the re-read found — chunks were cut by character count in front of a 512-token window, so 13.7 % of the published Japanese chunks and 18.5 % of the Chinese ones ran past the encoder limit and lost 21.7 % and 27.0 % of their tokens at embed time; a 184-character English licence footer sat inside 97.1 % of the Japanese chunks and 92.9 % of the Chinese ones; and the shipped index type returned 45.4–47.1 % of the exact top-10, costing 7.8 MRR in Japanese and 10.7 in English.
• What the rebuild did — and did not do — retrieval against the previously published artifact moved from 26.6 to 43.3 MRR in Japanese and 26.2 to 47.5 in Chinese, on a control question set generated from the old extractor’s own output. The components are not separated and the rebuild is not uniformly better: indexed-article coverage falls in two of three languages.
• Review and scope — three rounds of adversarial review are reported at equal prominence with the results; round three found a defect in the release that rounds one and two had both passed. The paper reports what happened to three specific artifacts and claims nothing about indexes elsewhere: the retrieval numbers come from synthetic questions with no human relevance judgements, and every reviewer was a model instance commissioned by the author.
  zenodo.org/records/22782085 · the rebuilt indexes on Hugging Face

Executable Preregistration & Act-Level Governance (2026.8)

Two papers from late August 2026: one registers a study whose protocol its own instrument enforces, the other states the act-locus argument the governance stack is built on.

• Preregistration as an Executable Contract — a machine-enforced protocol for a paired N=800 evaluation of an LLM-agent assurance plane. The run is authorized once under a reserved nonce and the analysis path is fixed in advance, so several classes of protocol deviation are refused by the instrument rather than discouraged by the document. At publication the study is registered but not executed — what is verified is the instrument, not a result.
  zenodo.org/records/22138712
• Empty Text, Normed Acts — why pattern enumeration loses to attack variation, and why the mediation point that survives is not the text but the typed act at the boundary, checked against an explicit authorization norm.
  zenodo.org/records/22139691

Measurement & Governance Studies (2026.8)

Four empirical papers from August 2026 — pre-registered or adversarially reviewed, each shipping with a public study-artifact repository on GitHub.

• Measuring Machine Habitus & Machine Doxa — a pre-registered pair: the dispositional structure of a model population, and the consensus layer where all models agree.
  zenodo.org/records/21982393 · zenodo.org/records/21982857
• The Reflective Budget Governor — equilibrium-based early stopping for iterative refinement; reduces reflection budgets by 33–37% in the paper's evaluation, with a reference implementation.
  zenodo.org/records/22031367
• A Guardrail That Failed Quietly — a production injection defense that degraded silently, found and corrected through the project's own governance loop.
  zenodo.org/records/21903321
• Quantization Regime and Governed Routing — a 2×2 study of QAT Q4_0 versus PTQ Q4_K_M under a fixed governance stack.
  zenodo.org/records/21903706

The Theory Series (2026.8)

Six formal papers supplying the theorem layer beneath the MOBIUS architecture. Every claim is graded, every paper states its break conditions, and the T1 core is machine-checked in Lean 4.

• T1 — Termination by Novelty: anti-regress guards, with a machine-checked core — zenodo.org/records/21837211
• T2 — Sequential, Never Layered: why governance overlays must not stack — zenodo.org/records/21837213
• T3 — Why Reflection Should Be Shallow: finite optimal depth and the rumination regime — zenodo.org/records/21837215
• T4 — The Crossover Theorem: capacity crossovers and axis-blind gate impossibility — zenodo.org/records/21837217
• T5 — The Exact Perimeter of Semantic Energy: what I(Y;Z) does and does not license — zenodo.org/records/21837223
• T6 — What Two Gates Buy You: abstention cascades and log non-identifiability — zenodo.org/records/21837226