MMV: A Governed, Local-First Language-Model Runtime with an Honest-Measurement Culture — Systems Paper — Route-First Answer Entitlement, a Pre-Output Metacognitive Organ, and Release Lines That Refuse to Inherit Each Other's Claims

MMV is an open-source (AGPL-3.0) runtime that wraps local and hosted language models in a governance layer whose core primitive is not answer quality but answer entitlement: before any content is generated, a route decision classifies what kind of move the turn warrants — answer, ask, verify, dateboundanswer, reanchor, or abstain — and the route, not the model's fluency, is authoritative. This paper is the citable systems description of the public artifact (mobius-style/mmv). We describe (i) the route-first architecture and its protocol authority stack (L0), which exists in human-readable, machine-readable, and compact deployable forms with an explicit documentation-authority map; (ii) the pre-output metacognitive organ (QK): a 42-item question catalogue fired selectively by an empirically calibrated policy, recently extended with a capacity-conditioned selection mechanism that ships inert by default so the frozen baseline is bit-identical; (iii) the three release lines — Small (9B-class engine), Medium (12B-class), Large (120B-class route-transformer stack) — under a discipline in which each line carries its own evidence floor and is forbidden to inherit a sibling's quality claims; and (iv) the evaluation culture that we consider the system's most transferable property: primary evidence is frozen before analysis, pseudo-environment numbers were publicly corrected by real-interface measurement (a drop from 26/30-grade scores to 18/33, later recovered to 31/33 by an engine fix that the bad number forced), development phases are closed by saturation signals rather than by feature appetite, and every public benchmark is labeled candidate-grade until independently reproduced. The runtime's governed behaviors are exercised publicly by a one-paste trial overlay; its dose–response properties across model scales are measured in a companion paper. MMV's thesis, in one line: for deployed assistants, what a model is entitled to say is a more tractable and more portable control surface than how well it says it. AI co-observer: Claude Fable 5 (Anthropic) — working method only; the registered author is the human author alone.