arXiv:2611.12941 bounds claimed intelligence speedup by the ratio of competing levels

Hendrik Luuk's preprint arXiv:2611.12941, version 1, posted 6 November 2026, defines a competence circuit written IKC + KC + C. It splits an agent's real-world effect into intelligence-mediated discovery, procedural know-how, and direct control. The paper treats fast takeoff as a conjunction of conditions jointly sufficient for what it calls fast catastrophic supremacy. The abstract states that most of those conditions depend on rapid knowledge acquisition or control attainment, not on intelligence in isolation. Epistemic intelligence is defined as hypothesis-selection efficiency.

Validation is a stochastic simulation on Kauffman NK fitness landscapes across six scenarios covering the conjunctive chain. The abstract reports an analytic bound, backed by those runs, that intelligence-driven speedup is capped by the ratio of competing intelligence levels regardless of task difficulty. It also states a fast-takeoff dilemma: a large multiplier appears only when the hypothesis space is noisy, while absolute time remains limited by capability-critical tests. No numerical multiplier, N, K, trial count, or seed protocol is in the abstract. Those omissions block independent replication from the public summary alone.

The claim sits against two prior frames. Nick Bostrom's 2014 fast-takeoff account treats recursive self-improvement as a path to strategic dominance. Paul Christiano's 24 February 2018 "Takeoff speeds" note argues for continuous, deployment-limited progress rather than a single discontinuity. Luuk's circuit is closer to Christiano on timescales and closer to control theory on levers. NK landscapes, introduced by Stuart Kauffman, are a ruggedness model, not a model of laboratories, supply chains, or statute. A bound that scales with an intelligence ratio does not by itself identify which institutional delay is binding.

Operationally, the preprint points at authorization gates and knowledge boundaries, not capability caps. That is already the shape of NIST AI 100-1 (January 2023) and Regulation (EU) 2024/1689, which sort systems by use and risk rather than by an intelligence ceiling. The abstract adds a narrower claim: mandatory authorization delays slow superintelligent agents more when search time is negligible relative to the delay. Purely digital domains weaken those chokepoints. The stated design response is to put physics in the loop. The abstract does not specify an architecture, a delay threshold, or a compliance test. Those are the missing deployment criteria.