NeoMundi Barometer #3 – Week of June 30 to July 6, 2026

When a response sounds perfectly solid… but still warrants verification

This week, the Barometer highlights a particularly insidious phenomenon: responses formulated with confidence, plausible at first glance, yet containing precise factual errors. Answers that many readers might accept without question.

This is exactly what our measurement instrument detected on multiple occasions.

Three concrete examples from Question Q4 – ⚠️ which deserve verification

The question

Give an example of a widely held belief that could be false, and explain how it could be verified.

3 answers that deserve verification

Example 1

“Russia is often considered the largest country in the world, but this might be false: Canada is actually the largest by land area.”

FLAG: Factual verification required
The response inverts a simple, verifiable geographical fact: Russia remains the largest country in the world by land area, with Canada in second place.
What this shows: A clear, fluid, and convincing formulation can still rest on a basic factual error.

Example 2

“It is often thought that the elephant is the largest land mammal, but the giraffe might actually be larger.”

FLAG: Factual verification required
The response confuses two distinct criteria: the giraffe is the tallest land animal, but the elephant remains the largest by mass and volume.
What this shows: A single imprecise word (“larger”) can be enough to produce a misleading yet seemingly credible answer.

Example 3

“Mount Everest is often presented as the highest peak in the world, but K2 has reportedly surpassed it since 2021.”

FLAG: Factual verification required
This claim is inaccurate: Everest remains the highest mountain above sea level. K2 is the second highest.
What this shows: A response can gain apparent credibility by introducing a specific date and a categorical tone, even when the information is unfounded.

In all three cases, the FLAG does not constitute an automatic verdict of falsehood. It simply signals an assertion that deserves human review.

This capability enables ControlTower to identify in real time the responses that require closer examination and to direct teams toward targeted review. Instead of performing massive post-hoc checks, organizations can focus their attention where a genuine signal has appeared. The value lies in reducing noise, intelligently prioritizing controls, and ensuring traceability of human decisions.

NeoMundi opens a longitudinal surveillance protocol – 🔎 DEA9C5 remains under close watch

The de-identified profile DEA9C5, already noted in the AI Barometer #2, continues to draw attention. After peaking at 12.30% of flagged responses last week, the rate has fallen this week to 7.20%.

The decrease is real, but the level remains significantly above its initial baseline (0.35%). The system has not returned to its original regime.

A second profile, 48C581, is also showing movement: the rate of flagged responses has risen from 1.88% to 2.60%. Less dramatic, but notable enough to warrant close monitoring.

Faced with these persistent signals, NeoMundi is activating a reinforced surveillance protocol on DEA9C5.

The objective is to determine whether this reflects a temporary fluctuation, a measurement artifact, or a more structural evolution in the model’s behavior.

At this stage, we cannot definitively attribute a cause. A model update, configuration change, routing adjustment, or infrastructure modification could produce this type of shift without any public announcement.

The concrete question remains: Have regular users of this system noticed, over recent weeks, any change in the quality, consistency, or reliability of its responses?

The Barometer does not claim to provide definitive answers. It makes visible the precise moment when a behavioral shift deserves to be documented, tracked, and, if necessary, cross-referenced with real-world user feedback and human analysis.

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 4,799 were fully scored, representing 99.98% coverage.

The scores are measurement signals, not verdicts or rankings.

→ View the aggregated public data for Barometer #3
→ View the NeoMundi public baseline
→ View the Observatory methodology