The IRAM-Ω-Q papers build a simulation framework for studying how adaptive agents regulate uncertainty under noise, perturbation, and changing control orderings.
Introduces the IRAM-Ω-Q framework: a transparent computational model for studying how artificial agents regulate uncertainty, disturbance, and instability over time.
Paper 1 · arXiv:2603.16020
Studies whether regulatory demand depends on the trajectory by which an artificial agent reaches a target state. Using a continuous target-entropy ramp, the paper shows hysteresis in adaptive gain and compares regulation-first and disturbance-first control orderings.
Paper 2 · arXiv:2606.30975
Studies switching between regulatory modes and asks whether intermittent anticipatory control can reduce average regulatory effort through carryover effects. Intermittent regulation is not equivalent to a weaker continuous version of itself: switching carries its own measurable cost.
Paper 3 · arXiv:2607.17432
Explores how different sources of disturbance, including internal and induced noise, affect regulation, stability, and recovery from perturbation.
Can an artificial agent become more agentic when regulation increases the window between impulse formation and action commitment?
Critical Attention Systems is preparing a research book on attention, regulation, stability, and artificial agents. The book introduces the IRAM-Ω-Q research program in clear, accessible language, connecting simulation-based models of uncertainty regulation with broader questions in AI, adaptive control, and contemplative science. It examines how artificial agents may stabilize attention, recover from disturbance, and regulate internal uncertainty over time. An early-access edition is planned through Critical Attention Systems Publications.
A separate line of work re-tests the framework's mechanisms on substrates with no quantum-like structure, using code executed on real infrastructure rather than simulation.
Tests the framework on a real language-model tool-use agent. When a governance layer's own interventions carry a cost, a cost-blind governor drives the agent it supervises into a self-reinforcing collapse; a governor that tracks the consequences of its own corrections and has the authority to do less recovers most of the lost performance, and works best at an intermediate degree of self-restraint.
Cross-substrate study · CASAgentGovernor
The goal is to build inspectable computational models and publish analyses that can be tested, criticized, and refined.
The publications do not claim to solve consciousness or general intelligence. They develop controlled simulation tools for studying regulation, uncertainty, timing, disturbance, and stability in artificial agents.
© Veronique Ziegler · Critical Attention Systems