Composing and Hardening a Single Decision Token: A Pre-Registered Characterization of Inference-Time Channel Fusion, Its Envelope, and a Cheap Noise-Robust Head
Year: 2026
Type: preprint
Abstract: Composing and Hardening a Single Decision Token: A Pre-Registered Characterization of Inference-Time Channel Fusion, Its Envelope, and a Cheap Noise-Robust Head Randolph James Ferlic, M.D., and Kimberly Kate Ferlic — Fieldstone Analytics, LLC Correspondence: [email protected] Preprint · Zenodo DOI: 10.5281/zenodo.22046713 · CC-BY 4.0 Abstract A class-discriminant codebook reduces each window of a multivariate sensor stream to a single eight-bit discrete token — features projected onto a shrinkage-regularized discriminant subspace, quantized to one of K = 256 centroids, and read by a lightweight lookup head — trained once and deployed frozen. The token is cheap enough for an always-on microcontroller, but its behavior once deployed and composed has not been mapped. Under a pre-registration discipline with honest negatives, we ask three deployment questions — when composing the token with an independent cheap channel at inference helps, when that help saturates, and how the token is hardened against noise — on three public benchmarks (Case Western Reserve University bearing vibration, UCI Human Activity Recognition, and the ROSMA da Vinci Research Kit surgical kinematics). We report that fusing the single token with an independent co-channel (a gain-invariant log-polar transform view or a diverse-model panel) by an equal-weight, parameter-free product rule is a no-harm, drift-robust composition that generalizes beyond vibration to non-oscillatory, per-individual cross-surgeon drift, recovering +0.057 macro-AUC (95% CI [0.019, 0.195]) where the token alone loses accuracy — and that the correct fusion partner under drift is the single token, not a sharper multi-token ladder, whose peaked posterior degrades the combination. The benefit is confined to the regime in which neither view dominates: it does not compound with additional channels and does not exceed a heavier full-feature model, so its value is edge-cost, not a raw-accuracy ceiling. Finally, additive input noise produces a cliff rather than a graceful decline, best repaired at the read-out by re-estimating only the lookup head on noise-augmented data — a closed-form, gradient-free change with the encoder and codebook frozen — recovering a large fraction of the loss (CWRU +0.13 macro-AUC, 95% CI [0.10, 0.16]), while silencing unreliable codebook cells does not. We do not propose a new fusion operator; we map where the established principles of decision-level fusion, ensembling, and robust estimation pay off on an extreme-compression single-token substrate — and, as importantly, where they do not. Every headline is pre-registered with outcome bands fixed before the data are examined, and every outcome, including those that did not meet their bands, is reported and reproducible on public data. Highlights • Fusing the single decision token with one independent cheap channel (a gain-invariant log-polar view or a diverse-model panel) by an equal-weight, parameter-free product rule is a no-harm, drift-robust composition — it does not hurt where there is no drift tax and recovers a portion of the loss where there is, generalizing beyond vibration to non-oscillatory cross-surgeon drift (+0.057 macro-AUC, 95% CI [0.019, 0.195]). • Counterintuitively, the correct fusion partner under drift is the single token, not a more-informative multi-token ladder: the ladder's sharper posterior, once drift-miscalibrated, dominates and degrades the equal-weight combination — yielding a threat-based composition map (the multi-token ladder for a lost or corrupted sensor without drift; the single-token-plus-channel combination for operating-condition drift). • The benefit is bounded: it is super-additive only where neither view dominates, does not compound with additional channels, and does not beat a heavier full-feature model — the value is edge-cost (near-model decisions from cheap views at a fraction of the representation and compute), not a raw-accuracy ceiling. • Additive input noise produces a cliff, best repaired at the read-out: re-estimating only the lookup head on noise-augmented data — closed-form, gradient-free, encoder and codebook frozen — recovers a large fraction of the loss (CWRU +0.13 macro-AUC, 95% CI [0.10, 0.16]); silencing unreliable cells and re-weighting projection directions do not. • No new operator is claimed: the contribution is a pre-registered map of where decision-level fusion, ensembling, and robust estimation pay off on an extreme-compression single-token substrate — with all outcomes, including those that missed their bands, reported verbatim. What this record contains • The manuscript (PDF). • A reproducibility archive (`PAPER_38_ZENODO_ARCHIVE.zip`): deterministic runners for every experiment (the composition map and the single-token-versus-multi-token pairing, the non-oscillatory cross-surgeon generality, the saturation/envelope study, the noise cliff and the noise-augmented robust-reader head, and the bootstrap-significance analyses), the per-run and per-seed JSON result summaries behind every figure, and the figures. All datasets are public (the Case Western Reserve University Bearing Data Center; the UCI Machine Learning Repository Human Activity Recognition dataset; and the ROSMA da Vinci Research Kit surgical-kinematics dataset). No raw subject data is redistributed; loaders fetch the public sources. Cite as R. J. Ferlic and K. K. Ferlic, "Composing and hardening a single decision token: a pre-registered characterization of inference-time channel fusion, its envelope, and a cheap noise-robust head," Zenodo, 2026, doi: 10.5281/zenodo.22046713. 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 described herein — including the inference-time channel-fusion, composition-map, and noise-augmented-head embodiments — are the subject of pending U.S. patent applications. Companion deposits Part of the single-token class-discriminant codebook family on Zenodo (community: spiral-domain-encoder-campaign), which includes the deterministic multi-token token-ladder deposit (doi: 10.5281/zenodo.22003179) and the self-validating free confirmation channel for on-device drift adaptation (doi: 10.5281/zenodo.21644285). Keywords decision-level fusion, product rule, deep ensembles, distribution shift, test-time robustness, class-discriminant codebook, single-token encoder, edge inference, condition monitoring, surgical kinematics, noise robustness, pre-registration.
Language: English
Subfield: Control and Systems Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7203899746