Public Evidence Archive for "Reliability-guided selective specialist routing for resource-aware hybrid classification"

Versioned public evidence archive for the RGSS framework. Version 2 retains the historical August 2026 archive unchanged and adds corrected tabular replication, small-data runtime records, paired predictions and the Fashion-MNIST seed-101 audit supporting the revised manuscript, "Reliability-guided selective specialist routing for resource-aware hybrid classification". The additional package contains nine-seed Dry Bean, Obesity and Drug Consumption/cannabis evidence; five-seed Wine, Balance Scale and Vehicle predictions; resident CPU inference measurements; and an A1/A5 scope audit. All included valid seeds, including unfavorable results, are retained. It contains no implementation source code, trained model weights, input features or facial images. Runtime boundaries: resident inference includes preprocessing through final decisions, but excludes disk/model loading. Reused Fashion-MNIST checkpoints do not establish matched fresh training-time savings. Earlier small-data final-fit training comparisons affected by a full-AUX seed mismatch are retained only for provenance, not as matched speedup evidence. Per-test-set diagnostics and repeated-holdout comparisons must not be pooled as independent datasets. See PUBLIC_VERSION_NOTES.txt and the protocols for interpretation, limitations and SHA-256 verification. Older README statements that the additional evidence has not yet been deposited describe pre-deposit status and are superseded by publication of this version. The historical archive retains its original title; no journal acceptance is implied.