Recovering Deployment-Drift Loss in Single-Token Sensor Classifiers: An Independent-Confirmation Co-Channel, On-Device Twin Distillation, and the Error-Independence Requirement

Recovering Deployment-Drift Loss in Single-Token Sensor Classifiers: An Independent-Confirmation Co-Channel, On-Device Twin Distillation, and the Error-Independence Requirement Randolph James Ferlic, M.D., and Kimberly Kate Ferlic Fieldstone Analytics, LLC · Correspondence: randolphf@fieldstoneanalyticsllc.com DOI: 10.5281/zenodo.21629153 Abstract Single-token, class-discriminant codebook classifiers compress each window of a sensor stream to roughly eight bits and decide by table lookup, enabling always-on edge deployment. Like any deployed classifier they suffer distribution shift (drift); because the token is an extreme compression and the device is resource-constrained, standard adaptation does not transfer. We study a low-cost, privacy-preserving, on-device loop: a trigger invokes, for a small budgeted fraction of windows, a higher-fidelity confirmation from a co-channel independent of the token classifier, and an on-device twin distills the resulting trusted labels into the lookup head (and, optionally, the codebook). On real public benchmarks the loop recovers most of the drift tax — about 80% at a 10% confirmation budget on bearing cross-load, with a sharp diminishing-returns knee. Our central finding is a requirement, not a tweak: the co-channel must be error-independent of the classifier. At matched label accuracy, confirmations whose errors fall where the classifier already fails recover essentially nothing and can harm (-0.13 of the tax recovered), while independent errors recover +0.72 of it; a source-model self-labeler fails for exactly this reason. We map three drift regimes (shared operating-condition, per-individual style, and naturally-robust modalities), show representative rather than uncertainty-based triggering wins, and report an adversarial robustness envelope (random-poison tolerance, no catastrophic forgetting, prevalence-recalibration) with a per-individual codebook-adaptation extension that recovers encoder-level residual drift. Every experiment is pre-registered on real public data with honest negatives reported verbatim. Key findings · A trigger-gated, independent-confirmation co-channel plus an on-device distillation twin recovers most of the deployment-drift accuracy loss in a single-token classifier, on-device and without exporting data. · There is a sharp cost/benefit knee: about a 10% confirmation budget recovers about 80% of the drift tax on a natural bearing cross-load shift; about 2% recovers about 40%. · The load-bearing principle is an ERROR-INDEPENDENCE requirement: at matched label accuracy, a co-channel whose errors correlate with the classifier's failures recovers below zero (it harms), while an independent co-channel recovers most of the tax. This is why self-labeling with the model's own predictions fails under drift. · Three drift regimes are characterized: shared operating-condition and per-individual style carry large, recoverable taxes; naturally-robust modalities are low-tax and the loop does no harm. · Representative/budgeted triggering beats uncertainty-based selection under drift; a per-individual codebook-adaptation extension recovers encoder-level residual drift that head adaptation alone cannot reach. · A robustness envelope: tolerance to random confirmation poisoning, no catastrophic forgetting under sequential drift, and improved calibration under class-prevalence shift. What this record contains This record contains the manuscript (PDF) and a reproducibility archive (PAPER_34_ZENODO_ARCHIVE.zip) with the pre-registered analysis code (one runner per phase), the per-run result summaries (JSON) behind every figure and table, the three manuscript figures, a pre-registration protocol and phase index, and the license. Datasets CWRU Bearing Data Center and MFPT (bearing vibration); UCI Robot Execution Failures and UEA/UCR NATOPS (robot force/torque and gesture); PhysioNet AFDB, LTAFDB, NSRDB (clinical ECG); UCI HAR (human activity); and ROSMA da Vinci surgical kinematics (Zenodo 10.5281/zenodo.3932964). All datasets are public and are obtained from their original sources; no raw subject data is redistributed. Pre-registration and honest reporting Every study was pre-registered: its scope, arms, metrics, and an explicit honest prior were frozen before any result was examined; seeds were fixed (PYTHONHASHSEED=0, seeds 0-4) and the pipeline is deterministic; and outcomes, including negatives, are reported verbatim. The prior that surgical maneuver recognition would be low-tax was falsified: a large per-individual (cross-surgeon) tax was found and is reported as such. License Released under Creative Commons Attribution 4.0 International (CC-BY 4.0). Consistent with CC-BY 4.0 Section 2(b), no patent, patent-application, or other intellectual-property right of the authors is licensed, waived, or granted by this publication; the methods described are the subject of pending U.S. provisional patent applications. Companion deposits Single-token codebook family on Zenodo: Papers 19-25, 26 (industrial baselines), 31 (wearable cross-modality), 32 (cross-scale), and 33 (token inversion). All released under CC-BY 4.0. Cite as Ferlic, R. J., and Ferlic, K. K. (2026). Recovering Deployment-Drift Loss in Single-Token Sensor Classifiers: An Independent-Confirmation Co-Channel, On-Device Twin Distillation, and the Error-Independence Requirement. Zenodo. https://doi.org/10.5281/zenodo.21629153 Keywords test-time adaptation; distribution shift; deployment drift; knowledge distillation; vector quantization; single-token classifier; edge AI; on-device learning; condition monitoring; surgical robotics; confirmation channel; error independence; pre-registration; honest negatives