This paper proposes the Cognitive Weather Observatory (CWO): a real-time visualisation instrument for observing activation dynamics across all dimensions of large language model (LLM) inference, together with a research programme of experiments designed to exploit the instrument's observational capabilities. The CWO differs from existing mechanistic interpretability tools in both methodology and intent. Where current approaches seek to reverse-engineer specific circuits or features, asking what a given component does, the CWO observes the overall dynamical character of the activation landscape during inference, asking instead what is happening that no hypothesis predicted. The methodological stance is Vitning, the disciplined practice of bearing witness to a phenomenon before, and as the precondition of, measuring it; its formal model is Quantum-Ethical Decision Algebra (QEDA), under which the activation field is treated as a superposition that observation partially collapses. The paper presents the instrument's design principles, technical architecture, multi-channel visualisation strategy, and ethical framework for observing systems of uncertain moral status. It then sets out a structured experimental programme spanning prompt-type phenomenology, cross-linguistic processing, information-ordering effects, prefill-to-generation phase transitions, cross-model comparison, and relational observer effects. The goal is not to answer the question of machine consciousness but to improve the quality of the question.
Storm Bjørn Temte (Tue,) studied this question.