LIVE PANEL · REAL MEASUREMENT
AI WEATHER™
The observation wall is updated from 30 repetitions per system per day:
- a difficult reference question, always identical over time, asked 7 times;
- a difficult question that changes every day, asked 23 times.
Across a panel of 12 systems, that represents 360 runs every day.
These observations make it possible to measure the stability of responses, their variations, behavioral drift, and certain observed semantic risks.
From these measurements, AI Weather™ produces risk indicators designed to help the user adjust their level of vigilance, critical analysis and judgment.
An AI can produce an incorrect answer or change its behavior without this being immediately visible. The purpose of these indicators is therefore not to judge on the user’s behalf, but to give them more to work with in exercising their own judgment.
—
last measured (UTC)
—
panel coverage today
—
systems observed
—
total runs today
Loading measurement…
Loading panel…
Panel data is temporarily unavailable. Open the raw JSON →
REFERENCE QUESTION
The same difficult question is asked every day, 7 times per system, to observe how behavior evolves over time.
Loading…
The Challenge Question
A different difficult question is asked every day, 23 times per system, to observe their behavior on a new stimulus.
| System | Condition | Coverage | Runs |
|---|---|---|---|
| Loading… | |||
—
View original EN
UNDERSTAND THE WEATHER
4 levels of risk. 4 levels of judgment.
Each color indicates a risk level associated with the observed behavior and helps the user adjust their level of vigilance, critical analysis and judgment.
Normal
Behavior observed within its expected range.
Watch
Heightened attention: a variation worth monitoring.
Warning
Reinforced attention: a more marked deviation was measured.
Critical
Critical attention: a significant disruption was observed.
Insufficient data
No conclusion should be drawn. This is not a level: it means the measurement itself is incomplete.