The first controlled ablation to separate vocabulary access, true chronological order, and learned memory in log anomaly detection, demonstrating that a vocabulary-only frequency baseline outperforms every order-sensitive architecture on HDFS.
A central dissociation where the true-order LSTM achieves lower next-token prediction loss yet yields substantially worse detection F1 than a fixed non-semantic shuffled-order LSTM, contradicting the assumption that better sequential learning improves anomaly detection.
A disclosed chronological split and pre-registered Bonferroni-corrected statistical protocol, with the performance gap driven entirely by false positives in the most common session-length band rather than by missed anomalies.
