Auditing what an observability forecasting foundation model actually encodes, and separating what a probe can recover from what the model demonstrably uses. Accepted at NeurIPS 2026.
Oct 4, 2026