A remarkably stable consensus now governs public and expert discussion of the near AI future: capability scales with resources and with inference-time thinking; agentic systems will absorb the majority of digital labor within years; AI-driven research is entering a self-acceleration loop; and open-weight diffusion is commoditizing capability into a heterogeneous fleet of models of every size. We accept this consensus — including its well-documented tendency to understate how fast its own predictions arrive. Our position is that the consensus is half of a description. It specifies the engine of the coming decade and leaves the steering unspecified, because it treats governance as either a property frontier laboratories add to frontier models or a matter of external regulation. Neither posture reaches the actual deployment surface: fleets of heterogeneous, increasingly autonomous, partly self-modifying agents operated by parties who are not frontier laboratories. This paper states the MOBIUS project's position in five theses, each anchored in released, measured, or live-operated work rather than speculation: (1) governance is a dosed intervention whose effect is conditioned on model capacity, with a measured crossover near 4B parameters and a measured penalty for over-dosing at every non-reasoning scale up to 120B; (2) the binding constraint on recursive self-improvement is not capability but anti-regress — drift, criteria capture, and meta-deliberation without state update — and this constraint admits a control-form answer (bounded admissible transitions, saturation and loop detectors, multimodality recovery) that we operate in a live self-evolving system; (3) the scalable safety primitive for agent-speed economies is reversibility, not prior approval; (4) reflection is bounded and shallow, so "let the model think more" is not a governance substitute; (5) answers and actions need entitlement and provenance, not just accuracy. We locate the genuine intersections and the one genuine disagreement between these theses and the capability consensus, and we state falsifiable commitments by which the position can be tested. The paper is offered as the anchor record for the project's Zenodo deposits: the single place where the question "why does MOBIUS exist?" is answered in terms commensurable with the mainstream debate. AI co-observer: Claude Fable 5 (Anthropic) — working method only; the registered author is the human author alone.
Toeda Taiko (Mon,) studied this question.