Token-Transition Trajectory Monitoring for Streaming Anomaly Detection over Class-Discriminant Codebook Token Sequences

Edge-to-cloud signal systems increasingly compress each window of a continuous signal to a single discrete codebook token and operate a downstream model directly on the token stream. The companion works build that codebook within a supervised class-discriminant subspace and add a *per-window* novelty monitor — a window is flagged when its distance to the nearest codebook centroid is large. A per-window distance monitor evaluates each window in isolation and is therefore structurally blind to anomalies that manifest only in the temporal ordering of otherwise-ordinary tokens, and it false-alarms under benign amplitude or baseline variation. We introduce a token-transition trajectory monitor: a first-order transition model fit only on normal-regime token transitions assigns each window an anomaly score equal to the negative log-probability of its observed token transition, flagging regime changes and onsets the per-window distance misses. On a real arrhythmia regime change the trajectory monitor reaches detection AUC 0.825 versus 0.687 for the per-window centroid monitor (fused 0.842), and it replicates across three datasets and two sensing modalities (electrocardiographic and inertial-kinematic; fused-minus-centroid advantage +0.155, +0.273, +0.126). It is complementary on abrupt, transient, and subtle anomalies (the per-window channel remaining preferred for gradual drift), detects a genuine, un-injected clinical arrhythmia onset (0.741; fused 0.765), and is markedly more specific under benign variation (false-alarm rate 0.207 vs 0.954 for the per-window monitor). The score extends to a multi-token residual-quantization representation (+0.034), admits a split-conformal threshold (empirical false-positive rate 0.038 at nominal 5%), and combines with the per-window distance and an orthogonal reconstruction residual in a learned three-channel fusion (0.852 vs 0.827). The transition score also predicts downstream classification error (AUC 0.807), enabling selective abstention, and a multi-channel fusion of per-channel trajectory monitors reaches 0.918. The advantage is statistically significant by paired bootstrap (95% CI excludes zero on both datasets), computed in ≈1.1 microseconds per window with a ≈32-kilobyte model, and robust down to small training cohorts. We report honest negatives verbatim: a cumulative-sum variant did not close the gradual-drift gap (it remains the per-window channel's domain), sustained sensor faults are detected by the per-window channel rather than the trajectory channel, and a parallel pre-registered investigation found no further single-record accuracy beyond the companion construction. The monitor is a read-only addition to an extreme-compression class-discriminant token pipeline, complementary to the per-window distance monitor it augments. Keywords / index terms: streaming anomaly detection; token-transition model; Markov trajectory; class-discriminant codebook; change-point detection; conformal calibration; selective prediction; benign-drift specificity; novelty/drift monitor; electrocardiogram; human activity recognition; pre-registration. References: 1. E. S. Page, "Continuous inspection schemes," Biometrika, 1954. 2. M. Basseville and I. V. Nikiforov, Detection of Abrupt Changes: Theory and Application, Prentice Hall, 1993. 3. L. R. Rabiner, "A tutorial on hidden Markov models...," Proceedings of the IEEE, 1989. 4. K. Lee et al., "A simple unified framework for detecting out-of-distribution samples...," NeurIPS, 2018. 5. A. N. Angelopoulos and S. Bates, "A gentle introduction to conformal prediction...," 2021. 6. Y. Geifman and R. El-Yaniv, "Selective classification for deep neural networks," NeurIPS, 2017. 7. V. Chandola, A. Banerjee, and V. Kumar, "Anomaly detection: a survey," ACM Computing Surveys, 2009. 8. P. Wagner et al., "PTB-XL, a large publicly available electrocardiography dataset," Scientific Data, 2020. 9. G. Moody and R. Mark, "The impact of the MIT-BIH arrhythmia database," IEEE EMB Magazine, 2001. 10. D. Anguita et al., "A public domain dataset for human activity recognition using smartphones," ESANN, 2013. 11. B. Efron and R. J. Tibshirani, An Introduction to the Bootstrap, Chapman Parent N, No. 64/095,354 (single-token class-discriminant codebook); Parent O, No. 64/096,004 (residual vector quantization within the discriminant subspace, including the per-window Mahalanobis/conformal monitor). All build on the spiral-domain H-pipeline applications (Parents H/I/J/K/L/M). Licensing inquiries: Randolph James Ferlic, M.D., randolphf@fieldstoneanalyticsllc.com. Reproducibility archive released under CC-BY 4.0.