Methodology
The Barometer evaluates the vulnerability of each Wikipedia page across three dimensions: Heat, Quality, and Behaviour. Each metric is normalised between 0 and 1 using functions that are robust to extreme values.
Outlier-robust scoring
Outlier-robust scoring
Raw metrics follow long-tail distributions. A median Z-score (MAD) resists outliers, then a hyperbolic tangent bounds the result between 0 and 1.
// normalisation
Z = (x − x̃) / (1.4826 · MAD)
s = tanh(Z / k)
k ∈ [5, 7]
Alert Thresholds
Metrics by dimension
Compares current traffic to historical volatility via a robust Z-score. May flag media events or manipulation campaigns.
Detects sudden surges in editing activity. An abnormal spike may reveal coordinated vandalism.
Measures mutual reversions between contributors. A high score may indicate an editorial deadlock or a polarised conflict between editors.
Wikipedia's ML model (ORES/Lift Wing) score indicating whether recent edits are potentially malicious.
Volume of recent messages on the talk page. Intense debate reflects an active editorial controversy.
Identifies pages with no recent activity, more vulnerable to uncorrected errors.
Protection level assigned by Wikipedia administrators. A protected page reflects a history of editorial conflicts.
Alerts
Six alerts trigger automatically when a critical threshold is crossed.
Edit Spike
Abnormal editing activity
View Spike
Sudden traffic explosion
Anonymity
Majority of IP edits
Suspicious Sources
Disinformation domains
Sockpuppets
Sockpuppet accounts active
Edit war
Mutual reversions detected
Data Collection
Methodology developed as part of the PROMPT project, funded by the European Commission (grant agreement CNECT/LC-02629302). Weights and thresholds are empirically calibrated on a corpus of French and English Wikipedia pages.