Stability, drift and a point of balance: exploratory research
NeoMundi is exploring a methodology for observing the behavior of artificial intelligence systems during execution and across repeated trials.
How can the stability of a response be characterized? What kinds of changes can be described as drift? Under what conditions do these variations become interpretable? These questions guide our research.
We are exploring, in particular, the idea of a point of balance between the development of a response, its coherence, and how it evolves in relation to the intended topic. This point of balance is a working hypothesis whose definition and usefulness remain to be tested across situations.
The hypothesis behind the instrument is that indicators calculated from observable outputs could help describe certain variations and guide their examination. We are studying their repeatability, sensitivity, and usefulness under documented conditions.
The methodology is developed through controlled observations, comparisons between texts and signals, and examination of the limitations of the measurement chain. Results must be interpreted alongside their protocol, configuration, and instrument version.
A signal alone does not establish an error, a hallucination, or general drift. The goal is to build a foundation of observations that can be discussed, tested, and improved; interpretation and decisions about use remain separate.