The Velocity Before the Plateau: Can a Maximum Admissible Rate of Frontier AI Growth Be Derived from Wisdom, Trust, and Governance Capacity?

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Frontier artificial intelligence is commonly governed through capability levels, evaluation thresholds, deployment gates, and risk categories. These instruments are necessary, but they are primarily state-oriented: they ask what a model can do at a given point in time. Recent frontier practice has made a second variable increasingly difficult to ignore. Anthropic has publicly argued that risk prevention may require pacing capability advancement so that safeguards have time to keep up, while OpenAI has reported temporarily slowing parts of model scaling in response to critical cyber-capability evidence. In parallel, empirical work on long-task horizons, AI research-and-development capability, and recursive self-improvement is beginning to measure not only what frontier systems can do but how rapidly the frontier itself may move. This paper asks whether that rate can be treated as a scientific object rather than only as a managerial or policy choice. The proposal develops a conditional dynamic model in which frontier capability is represented by a dimensionless index ι(t), while the capacities required to keep that capability admissible are represented through four non-compensatory channels: safety, control, evaluation, and governance. Building on the archetypal Wisdom Equation W = Iᶜ while preserving its original non-empirical status, the paper introduces a frontier-local operational projection C_F of consequence-consciousness and defines Wisdom Debt as D_W = (1 − C_F) ln ι. A finite Wisdom Reserve then permits temporary acceleration while making explicit that acceleration consumes margin when consequence-sensitive capacity does not grow at the same rate. Differentiating the debt relation yields a Wisdom Pacing Bound. Independently, barrier-style invariance conditions yield an engineering bound determined by the most restrictive protection channel, with extensions for latency, uncertainty, and exogenous requirement drift. The Maximum Admissible Frontier Velocity is defined as the minimum of the wisdom and engineering bounds. The resulting Trust Plateau is not a claim that intelligence must stop improving. It is a region in which additional capability can no longer justify an equal increase in autonomy, deployment scope, trust, or authority until consequence-handling capacity recovers. This reframes the apparent opposition between acceleration and plateau: recursive capability growth may continue while admissible authority plateaus. The paper offers propositions, illustrative simulations, measurable variables, falsifiers, and a calibration program. It does not claim a universal numerical speed limit, phenomenal machine consciousness, or proof that singularity or abundance scenarios are impossible. Its narrower hypothesis is that frontier capability velocity can be made conditionally auditable by tying it to finite rates of safety, control, evaluation, governance, reversibility, and consequence-aware response.