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
Reading causal self-attention as a probabilistic model surfaces a barrier constraint, a stability margin analogous to an SVM's, and the support tokens that sit on it. The training change is one log-barrier penalty added to cross-entropy.
Feb 25, 2026