The Wrist-PPG Tier-0 Gate: A Deep, Broad, and Adversarially Stress-Tested Map of a Frozen Single-Token Encoder as an Always-On Wearable Physiological Classifier and Anomaly Monitor

Abstract

The Wrist-PPG Tier-0 Gate: A Deep, Broad, and Adversarially Stress-Tested Map of a Frozen Single-Token Encoder as an Always-On Wearable Physiological Classifier and Anomaly Monitor Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: 10.5281/zenodo.23000837 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign · a characterization of previously filed and published methods; no new algorithmic subject matter is disclosed. Abstract The wrist wearable is the archetypal always-on, battery-bound sensor: a photoplethysmogram (PPG) — the optical pulse waveform behind every smartwatch heart-rate reading — is sampled continuously, and something must decide, second by second, what physiological state the wearer is in and whether it departs from their own normal. That perpetual sensing-and-triggering layer is the wearable instance of the Tier-0 tier we have argued the 2026 agentic edge leaves unowned. We study a frozen, deterministic, class-discriminant single-token encoder (a fixed 128-dimensional PPG micro-morphology front-end → a supervised linear-discriminant ⊕ principal-component subspace → a k-means codebook of at most 256 cells → an 8-bit token → a per-cell lookup-table decision; and, without labels, a per-entity k-means "self-twin" novelty detector) as that always-on wearable layer, and build a predictive, deployment-facing map of where it wins, pays, and fails across two public wrist-PPG corpora and five physiological channels through ~64 pre-registered results. The token is near-parity on few-class physiological classification (WESAD affect tax −0.019, 95% CI [−0.047, +0.008]; DaLiA rest/active +0.026), with a difficulty axis that is a low absolute ceiling for both models, not a token tax. A label-free self-twin flags stress (AUROC 0.877) and high-exertion (0.947) without labeled events, beats every "elevated heart rate" baseline, and matches a trained autoencoder deterministically. The central result is a pre-registered contrast: unlike the universal-extreme fall of our radar study, a PPG anomaly is entity-specific physiology, so per-entity self-twins are necessary (own > population on 100% of subjects) — and cheap (per-entity beats population with only 4–8 commissioning windows, under a minute of resting PPG). We characterize the full deployment surface — a confound-free motion test that overturns an apparent "crash" as a label-prior artifact, cross-device transfer, a cohort fairness tax, calibration, determinism, alarm fatigue, a within-session non-stationarity that inflates fixed-threshold false alarms while leaving AUROC robust, and a PPG biometric re-identifiability — and deliver cross-modality results: fusing the watch's own sensors (PPG + accelerometer) lifts the exertion gate to 0.985 at 1.3× energy, while a frozen cardiac codebook is modality-general but not modality-portable (PPG↔ECG does not transfer — re-commission per sensor). Two platform properties (one frozen encode serving four capabilities at zero interference; per-capability energy ~1/M) and a cross-modality predictive law (per-entity necessity iff the anomaly is entity-specific) round out an honest map. This is a characterization of previously described, filed methods; it discloses no new algorithmic subject matter, and the per-deployment / per-sensor selection of configuration is retained as trade secret. Highlights · Near-parity where few-class (WESAD affect −0.019 CI [−0.047, +0.008]; DaLiA rest/active +0.026); on PPG the difficulty axis is a low ceiling, not a token tax (8-class +0.009). · Label-free physiological gate — stress 0.877, exertion 0.947 with no labeled events; beats trivial heart-rate baselines; matches a trained autoencoder deterministically; a free signal-quality out-of-distribution gate (AUROC 1.00). · ⭐ Per-entity necessity confirmed and cheap — per-entity self-twins beat the population model on 100% of subjects (a pre-registered contrast to the universal-extreme fall), and do so with only 4–8 commissioning windows (< 1 min of resting PPG). · Full deployment surface — a confound-free motion correction (the apparent "crash" was label-shift), cross-device transfer, a cohort fairness tax (skin-tone unmeasured, flagged), calibration, determinism, alarm fatigue, non-stationarity (AUROC robust, fixed-threshold false-alarm rate inflates), and PPG biometric re-identifiability → emit the decision, keep the token ephemeral. · Cross-modality — multi-sensor fusion lifts the exertion gate to 0.985 at 1.3× energy (PPG + accelerometer is the sweet spot); a cardiac codebook is modality-general but not modality-portable (PPG↔ECG → re-commission per sensor); a foundation codebook is label-efficient at small budgets. · Platform — one frozen encode serves four capabilities at zero mutual interference; per-capability energy ~1/M; and PPG completes a cross-modality predictive law (per-entity necessity iff the anomaly is entity-specific) across four modalities. · Honest negatives — a low fine-classification ceiling, a corrected motion prediction, a bounded within-embedding composition, a pooled-few-shot negative (per-subject works), cross-modality non-portability, and biometric re-identifiability. What this record contains · Manuscript_Paper52.pdf — the manuscript (10 figures embedded, 77 references); and Manuscript_Paper52.docx, the editable source. · PAPER_52_ZENODO_ARCHIVE.zip — the reproducibility archive (md5 in ARCHIVE_MD5.txt): the frozen pre-registration; all runners (core accuracy; the gate + per-entity necessity; the skeptic/acquirer battery; exhaustive waves A/B; the tier-3 exhaustion battery; the tier-4 cross-capability/cross-modality sims — fusion, train-once many-heads, PPG↔ECG transfer, foundation; the tier-5 closing mine; composition and co-channel; the Modal fetch/featurize apps); the frozen token encoder and the fixed 128-dim PPG micro-morphology front-end; the per-experiment result records (JSON); the 10 figures; the manuscript source; and a README. Public datasets are not redistributed (fetched at run time); all paths and handles are scrubbed (PATH_TO_DATA/PATH_TO_SCRATCH, any cloud handle → MODAL_USER) and leak-scanned. Cite as R. J. Ferlic and K. K. Ferlic, "The wrist-PPG Tier-0 gate: a deep, broad, and adversarially stress-tested map of a frozen single-token encoder as an always-on wearable physiological classifier and anomaly monitor," Zenodo, 2026, doi: 10.5281/zenodo.23000837. License and patent notice Released under CC-BY 4.0. Consistent with that license, no patent or other intellectual-property right of the authors is licensed, waived, or conveyed by this deposit. This work characterizes previously described, filed, or published methods and discloses no new algorithmic subject matter; gradient boosting, k-means / vector and product quantization, Fisher discriminant analysis, PPG micro-morphology and heart-rate-variability front-ends, temperature scaling, autoencoder / isolation-forest / one-class anomaly detection, decision-level sensor fusion, and nearest-centroid novelty detection are established prior art, used only as tools. The methods characterized — the class-discriminant single-token codebook encoder and its nearest-centroid monitor, the multi-token / token-ladder and soft-readout mechanisms, the inference-time co-channel fusion, the foundation codebook, and the per-entity self-twin — are the subject of filed and pending U.S. patent applications held by the authors, including U.S. Provisional Application No. 64/095,354 (the encoder), the personalization / on-device-adaptation applications (priority U.S. Application No. 19/467,303 and its continuations), the multi-token / token-ladder application (No. 64/119,487), and the inference-time fusion application (No. 64/137,805). The observed per-entity-necessity result, the low-ceiling difficulty axis, the fairness and non-stationarity boundaries, and the cross-modality (PPG↔ECG) non-portability are properties of the frozen filed method, not new subject matter. The per-deployment / per-sensor selection of configuration and commissioning procedure is retained as a trade secret and is not disclosed here. © 2026 Fieldstone Analytics, LLC and the authors. Inquiries: randolphf@fieldstoneanalyticsllc.com. Companion deposits (spiral-domain-encoder-campaign) · The mmWave Micro-Doppler Tier-0 Gate (the companion radar map; its universal-extreme fall is the per-entity-necessity contrast established here): doi:10.5281/zenodo.22999908 · The Acoustic Tier-0 Gate (the always-on-audio map; the per-machine self-twin whose per-entity necessity is confirmed here for physiology): doi:10.5281/zenodo.22968134 · The Cheapest Digital Twin (the per-entity self-twin personalization): doi:10.5281/zenodo.22922678 · Where the Cheapest Token Wins, Pays, and Fails (the edge-AI Tier-0 map): doi:10.5281/zenodo.22945419 · The Configurable Bottleneck (the platform whose composition/multiplexing dials are exercised): doi:10.5281/zenodo.22884023 · Paying Down the Price of the Bottleneck (the label-efficiency / foundation-codebook study): doi:10.5281/zenodo.22866039 · The Price of the Bottleneck (the deployment/drift characterization): doi:10.5281/zenodo.22838118 · Label-free inference-time channel fusion (the multi-sensor fusion mechanism): doi:10.5281/zenodo.22046713 · Class-discriminant single-token codebook construction (the base encoder): doi:10.5281/zenodo.20788187 · Deterministic multi-token token ladder: doi:10.5281/zenodo.22003179 · The predictive reach of a decision token: doi:10.5281/zenodo.22736921 · Non-invertible but not anonymous (privacy): doi:10.5281/zenodo.22819210 · Unlinkable but not anonymous (privacy): doi:10.5281/zenodo.22838120 · The token as a bounded, threshold-free generative-edge cache key: doi:10.5281/zenodo.22148612 Keywords edge