Compose, Never Layer: MOBIUS INFINITY and the Entitlement–Reflection Orchestrator — Systems Paper — Sequential Composition of Answer Entitlement and Reflective Questioning, and the Measured Failure of the Layered Alternative
Toeda Taiko · Zenodo (CERN European Organization for Nuclear Research) · 2026
MOBIUS INFINITY is the synthesis capstone of the Möbius program: a thin sequential router (ERO — Entitlement–Reflection Orchestrator) that composes two independently governed, independently shipped systems behind one OpenAI-compatible endpoint. MMV answers the question "is answering warranted?" on its unmodified stack, and its route decision is authoritative; when the route is ask — the turn does not yet warrant an answer — ERO hands the turn to RQA, whose trade is the deeper question that advances it. Context flows one way (MMV → RQA), and the two systems are composed in sequence, never merged. The architectural rule in the system's name is not a stylistic preference but a measured constraint: composition experiments in the program's record (MMV "Condition I") documented that layering reflective content onto MMV's structural governance — injecting question-generation material into the entitlement stack itself — collapses restraint: the governed system's defining behavior, its willingness to withhold, degrades when reflection is fused into it rather than sequenced after it. INFINITY is therefore the program's constructive answer to its own refuted monolith: the founding vision of one integrated reflective stack survives as a federation — each member frozen, separately evidenced, separately citable — with orchestration kept thin enough to audit (route in, branch on ask, one-way context, no shared mutable state). We describe the composition contract, the deployment surface (fully local by default: one command serves an OpenAI-compatible API on consumer hardware with no cloud key; tiered reflection budgets keep the ask-branch latency of a local 12B host inside a streaming-masked 15–20 s), and the honesty boundaries: INFINITY makes no capability claim of its own — its claims are exactly the union of its members' published, candidate-grade claims plus one architectural theorem-by-measurement: compose, never layer. AI co-observer: Claude Fable 5 (Anthropic) — working method only; the registered author is the human author alone.