One profile shows a clear shift in zone, while the overall picture remains calm
This week, the overall average remains calm. However, one of the observed systems has changed its behavior and triggered our response verification control much more frequently compared to Barometer #1 last week.
Indeed, the anonymized LLM with the profile identifier DEA9C5 (see graph below) caught our attention due to a notable change: the percentage of responses marked as “needing further verification” rose significantly from 0.3% last week to 12.3% this week, a 41-fold increase.
Note: In this campaign, responses are reviewed by a second AI model tasked with detecting potentially weak factual claims. Its signal is meant to draw attention; it does not replace human verification nor constitute proof that a response is false.
This is the most significant observation from this week’s AI Barometer. It confirms the value of tracking changes over time: an AI may appear to function normally overall while evolving significantly in certain types of responses.
This is what we call a silent regime shift: a system’s behavior evolves without prior notice, often without the user noticing, and without any visible alert.
This does not mean the system has become bad, nor that all these responses are false. It simply shows that it no longer responds exactly as it did the previous week.
However, for any organization that uses AI to inform, assist teams, guide decisions, or respond to clients, it is crucial to detect behavioral changes before they lead to consequences.
This is why, every week, NeoMundi asks the same questions to the same systems. This allows us to detect what is actually changing, rather than relying on a single snapshot.
You can freely access all the public, aggregated, and de-identified data from this 2nd NeoMundi Barometer #2 on our GitHub repository.
The next observation will help determine whether this change disappears, stabilizes, or becomes confirmed.
NeoMundi Barometer · week 2 · 12 de-identified profiles · generated 29/06/2026 14:33
This cartography is generated from aggregated public data using a reproducible Python script, available here: view the cartography generator.
Methodology and public data
This Barometer tracks 12 de-identified profiles using 4 fixed questions, each repeated 100 times. The campaign includes 4,800 executions, of which 4,773 were fully scored, representing 99.44% coverage.
The scores are measurement signals, not verdicts or rankings.
→ View the aggregated public data for Barometer #2
→ View the NeoMundi public baseline
→ View the Observatory methodology