# CascadeLedger preprint reports detection of hidden information use in AI routing

_Published Friday, October 16, 2026 at 6:13 AM EDT · Science, AI · Latest · Tier 2 — Notable_

Researchers report an audit framework that detects when systems routing work among language models use information they were not entitled to read. Their CascadeLedger preprint reports detection rates of 0.9983 for serialized faults and 0.9982 for unlogged hidden-read faults, with a clean false-positive rate of 0.0000 across its evaluations.

The framework records reads and resource budgets, then changes hidden information while holding the visible execution history and policy randomness fixed to test whether routing improperly changes. Deterministic recomputation changed 0.3403 of corrupted quality-resource conclusions. The results cover frozen-cache, independently generated, and serving evaluations; online recording and offline auditing have separate overhead measurements.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2610.20668)

---
Canonical: https://techandbusiness.org/newswire/q1x5PTAwU6nfWcI1dPu6p0
Published: 2026-10-16T10:13:09.768Z
Story chronology: 2026-10-16T04:00:00.000Z
Retrieved: 2026-10-16T13:02:45.504Z
Publisher: Tech & Business (techandbusiness.org)