We present reptimeline, a Python library for tracking how discrete representations evolveduring neural network training. Unlike scalar logging tools (WandB, TensorBoard) that reportaggregate metrics, reptimeline tracks per-code lifecycle events: when concepts become distinguishable (births), when representations collapse (deaths), when relationships form betweenconcept pairs (connections), and where phase transitions occur in training dynamics. Thelibrary additionally discovers what each code element encodes—anti-correlated pairs, dependency chains, and three-way AND-gate interactions—without requiring prior ontological knowledge, with all discoveries subjected to multiple-comparison correction (Bonferroni or BenjaminiHochberg FDR). Causal verification uses bootstrap confidence intervals, permutation tests, andCohen’s d effect sizes. reptimeline ships three built-in extractors (VQ-VAE, SAE, FSQ), fivestatic and four interactive visualizations, JSON/CSV export, and a command-line interface.We validate on three architecturally distinct backends: a 32-bit binary autoencoder on MNIST(decoder determinism verified: 100%), sparse autoencoder features on Pythia-70M (8/16 features with finite KL selectivity, mean 26.8× L2 ratio with bootstrap 95% CI, plus 8 featureswith zero cross-activation attributable to SAE sparsity), and a 63-bit neurosymbolic projectionhead (crystallization event invisible to loss curves). reptimeline is the only tool that combineslifecycle tracking, bottom-up feature discovery, auto-labeling, causal verification, and theoryreconciliation in a single backend-agnostic package. Available via pip install reptimelineunder BUSL-1.1.
J. Arturo Ornelas Brand (Tue,) studied this question.