The NeoMundi AI Observatory is our public research program dedicated to the continuous, transparent and reproducible evaluation of generative AI systems.
We regularly publish thermodynamic cartographies and in-depth observations, including the NeoMundi AI Barometer, which tracks de-identified model profiles over time across multiple dimensions: stability, factual validity, informational density, factual-risk signals, semantic variation and cost efficiency. These publications reveal behavioral trends and silent regime changes that traditional one-shot benchmarks cannot capture.
All our work is built on open methodologies and publicly available data. Generation scripts, scoring protocols and raw datasets are released on GitHub, allowing researchers, developers and the wider AI community to audit, reproduce and build upon our observations.
All AI Barometers Observation Level: 🟡 Yellow This week, AI Barometer #7 highlights two complementary dimensions. On the one hand, the overall calm observed in AI Barometer #6 is confirmed: average semantic variation decreases, and the majority of profiles remain close to their usual regime. On the other hand, this return to calm is not […]
Fatima Ezzahrae Gouarab, Data Scientist, statistician and scientific contributor to the NeoMundi Research Observatory, is conducting work on the stability and reproducibility of the signals produced by NeoMundi ControlTower. DOI : https://doi.org/10.5281/zenodo.21499715 This first experimental phase is based on 680 generations, carried out with three model providers, two generation temperatures and several corpora covering factual
All AI Barometers Observation Level: 🟡 – This week, AI Barometer #6 explores two complementary dimensions. On the one hand, a form of volume variability: the same system can reach the same correct result while generating very different quantities of tokens. On the other hand, a return to calm after the peak observed the previous
NeoMundi Observatory ran a full runtime protocol on Kimi K3 (Moonshot): 1,200 valid executions – 120 situations repeated 10 times across 12 families (factuality, reasoning, governance, medical, legal, cybersecurity, business operations, and more). This is not a ranking.This is not a judgment.This is a runtime observation. The objective is to make visible what happens during
Reading note — 15 October 2026. This campaign took place during a period of maintenance of the NeoMundi measurement instrument. Observations are retained as originally published; comparisons with adjacent periods must account for the instrument configuration in use at the time. The narrative and summary card report different aggregates: 90.10% normal regime versus 81.01% mean
Reading note — 15 October 2026. July 2026 exploratory campaigns. The two protocols below are presented in separate sections and must not be pooled. Scores depend on the measurement chain, evaluator configuration and coverage used at the time. Agreement between AI judges is not independent proof of factual truth, and a constant indicator does not
Reading note — 15 October 2026. This article presents NeoMundi’s applied research approach. References to NIST, ISO and the AI Act provide context; they do not imply endorsement, certification or proof of regulatory compliance. Historical comparisons require checking protocols, evaluator versions and coverage. Companies no longer need to be told that artificial intelligence carries risks.
Reading note — 15 October 2026. This publication reports a June–July 2026 campaign. Figures describe signals recorded by the measurement chain used at the time, not verified error rates or a general assessment of model reliability. Cross-period comparisons require matching protocols, evaluator versions and coverage; they must not be carried over automatically to a later
Reading note — 15 October 2026. This is an underwriting perspective based on cases and campaigns cited at the time of writing. It does not establish actuarial validation, an insurance pricing method or a demonstrated reduction in claims. Historical signals require verification of their measurement chain; litigation references describe the original publication, not current procedural
Reading note — 15 October 2026. This publication reports a June–July 2026 campaign. Figures describe signals recorded by the measurement chain used at the time, not verified error rates or a general assessment of model reliability. Cross-period comparisons require matching protocols, evaluator versions and coverage; they must not be carried over automatically to a later
