A Self-Validating Free Confirmation Channel for No-Harm On-Device Drift Adaptation of Single-Token Classifiers

A Self-Validating Free Confirmation Channel for No-Harm On-Device Drift Adaptation of Single-Token Classifiers Randolph James Ferlic, M.D., and Kimberly Kate Ferlic — Fieldstone Analytics, LLC Correspondence: randolphf@fieldstoneanalyticsllc.com Preprint · Zenodo DOI: 10.5281/zenodo.21644285 · CC-BY 4.0 Abstract Single-token, class-discriminant codebook classifiers compress each window of a sensor stream to roughly one byte and decide by table lookup, enabling always-on edge deployment; like any deployed classifier they lose accuracy under distribution drift. A companion study established that this loss is recoverable on-device by distilling labels from a confirmation channel whose errors are independent of the classifier, but assumed a higher-fidelity physical co-channel and did not address how to obtain such a channel with no extra sensor, how to know on-device whether it is independent, or how to guarantee it never does harm. This paper closes those gaps. A purely software free channel — a diverse-model disagreement panel over the same input, or, for oscillatory signals, an analytic-signal log-polar view of the same window — supplies independent confirmations at no additional sensing cost, and a self-validating, no-harm-safe system is built around it: an unsupervised token-distribution-shift trigger operated at a deliberately high-recall point; a small ground-truth validation seed that estimates the channel's conditional error given the classifier is wrong and trusts the channel only where that error is low; a bounded-improvement criterion that guarantees no accuracy loss; a hard-label count-increment distillation of accepted labels into the head; and a channel-output quality gate. On real public benchmarks the system recovers 0.91–0.92 of a full-supervision oracle at 0.65–0.67 of the label cost on two large-tax families — industrial cross-load and surgical cross-operator drift — does no harm on a naturally-robust wearable family where it correctly abstains from spending labels, needs only about 25–50 seed labels, remains stable across a continual multi-shift sequence with no catastrophic forgetting, and resists correlated, systematic, and degenerate/poisoning channels. A head-to-head comparison shows the high-recall trigger is the correct operating point for a no-harm objective, because the precision-oriented alternative misses recoverable drift. The diverse-model and spiral channels are validated through the identical machinery, establishing an independent-channel genus with a broad and a narrow species. Every experiment is pre-registered with honest negatives reported verbatim. Highlights • A confirmation channel that recovers on-device drift loss can be obtained **in software, with no extra sensor** — from • model diversity or a domain transform of the same window. • Its error-independence is **certified on-device** from a small validation seed **before** it is trusted; otherwise the • system falls back to ground-truth labels. • A high-recall unsupervised trigger plus a bounded-improvement criterion give a **no-harm guarantee** that holds across • large-tax, low-tax, and adversarial conditions; the high-recall choice is shown to be correct for a no-harm objective. • Composed conditional-error and channel-output-degeneracy defenses make the system robust to correlated, systematic, and • targeted single-class poisoning channels. • Two channel species — a broad software panel and a narrow oscillatory log-polar view — pass the identical machinery. What this record contains • The manuscript (PDF). • A reproducibility archive (`PAPER_35_ZENODO_ARCHIVE.zip`): deterministic pre-registered runners, the per-run JSON • result summaries behind every figure and table, the data figures, and a pre-registration protocol with a per-study • index. All datasets are public (CWRU bearing; ROSMA da Vinci surgical kinematics, Zenodo 10.5281/zenodo.3932964; UCI • HAR); no raw subject data is redistributed. Cite as R. J. Ferlic and K. K. Ferlic, "A self-validating free confirmation channel for no-harm on-device drift adaptation of single-token classifiers," Zenodo, 2026, doi:10.5281/zenodo.21644285. License and patent notice Released under CC-BY 4.0. Consistent with Section 2(b) of that license, no patent, patent-application, or other intellectual-property right of the authors is licensed, waived, granted, or otherwise conveyed by this publication or by any reuse of it; the methods are the subject of pending U.S. provisional patent applications. Companion deposits Single-token codebook family on Zenodo, including the companion drift-adaptation study (DOI 10.5281/zenodo.21629153). Keywords test-time adaptation, distribution shift, confirmation channel, ensemble disagreement, self-training, no-harm guarantee, data poisoning, vector quantization, edge AI, condition monitoring, surgical robotics, pre-registration.