Track/Rail Algorithm v1.0.5: Signal-Augmented Expert Routing and Reproducible Evaluation

The Track/Rail Algorithm (TRA) is a heterogeneous
ensemble that learns to assign inputs to expert models using their
predictions and descriptors of local data geometry. This paper
updates its description to match the available version 1.0.5 source
and evaluates that implementation without modifying its learning
code. The method combines preprocessing, optional feature
selection, bootstrap or cluster-based expert training, a supervised
router, and an optional correction model. Source inspection shows
that several advertised capabilities require qualification: routing
entropy is inactive in the current feature construction; top-k
selection, capacity enforcement, and a load-balancing training
objective are absent from the active paths. A fresh study attempts
420 model–fold evaluations on four classification and two regression datasets using repeated five-fold cross-validation, with raw
predictions, failures, and source hashes retained. Results reveal
configuration-dependent accuracy and computational overhead,
while focused counterexamples identify inconsistencies between
label prediction and probability aggregation and an expert-index
error in hard routing. Historical benchmark files are audited
separately because their coverage and reported values do not
fully support the earlier manuscript. The study provides an
implementation-faithful specification and reproducible evidence,
while avoiding claims of universal superiority, verified drift
adaptation, or benefits from mechanisms that are not active in
this release