Deduplicating Correlated Evidence Breaks Calibration: A Shrinkage Correction for Quorum-Validated Multi-Agent Alerts
Removing correlated evidence creates the small-sample regime that breaks the naive confidence estimator; a shrinkage estimator makes calibration nearly invariant to the contamination rate. Expected calibration error: ECE 0.334 → 0.128. Files are under restricted access. Metadata, authorship and date are public; access to the full text is granted on request.
Authors
- Guilherme Röpke Schneider
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-08-14
- DOI
- https://doi.org/10.5281/zenodo.21926085
- Primary Topic
- Adversarial Robustness in Machine Learning
- Type
- preprint