The mmWave Micro-Doppler Tier-0 Gate: A Deep, Broad, and Adversarially Stress-Tested Map of a Frozen Single-Token Radar Activity Classifier and Fall Monitor

The mmWave Micro-Doppler Tier-0 Gate: A Deep, Broad, and Adversarially Stress-Tested Map of a Frozen Single-Token Radar Activity Classifier and Fall Monitor Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: [reserved at deposit] · CC-BY 4.0 · Community: spiral-domain-encoder-campaign Abstract Radar is the privacy-preserving always-on sensor: an FMCW or mmWave device senses motion continuously, through the dark and through clothing, without ever forming an image of a face or a room — which is exactly why it is the sensor of choice for in-home fall detection, elder-care activity monitoring, in-cabin automotive sensing, and contactless vital signs. That continuous sensing-and-triggering layer is the radar instance of the Tier-0 tier we have argued the 2026 agentic edge leaves unowned: an energy budget dominated not by the occasional expensive inference but by the perpetual cheap one. We study a frozen, deterministic, class-discriminant single-token encoder (a fixed 128-dimensional micro-Doppler 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 radar layer. The token's marginal cost is ~12,000 operations and a measured 20 microseconds per decision (49,000 decisions/second) atop a shared ~17-million-operation micro-Doppler front-end. Rather than merely demonstrate that the encoder "works," we build and adversarially stress a predictive, deployment-facing map of where it wins, pays, and fails across two public corpora and four radars — the University of Glasgow 5.8 GHz FMCW set (6 activities including falling, 106 subjects, spanning a laboratory, a common room, and two real elder-care facilities) and the CI4R cross-frequency set (11 activities recorded simultaneously on co-located 24 GHz, 77 GHz, and 10 GHz ultra-wideband radars) — through ~40 pre-registered results. On activity classification, the token is near-parity with a strong gradient-boosted model where the task is few-class (spoken-scale activities +0.022; a coarse grouping +0.006) — and this near-parity survives the strictest possible split (leave-one-collection-campaign-out, +0.026; a permutation null collapses it to chance) — while paying an honest, priced tax on many fine confusable gaits (+0.06 to +0.08). On label-free fall monitoring, a self-twin flags falls without any labeled falls at mean AUROC 0.886, is best-in-class at the safety-critical fall (per-class recall 0.959, exceeding the full model), beats every trivial "big/fast motion = fall" physics baseline, and — the map's central dissociation — improves when pooled into a population model (0.922 > per-subject 0.886), because a fall is a universal extreme event and therefore, unlike the per-machine and per-patient anomalies of our prior work, needs no per-person commissioning. We characterize the full deployment surface — a bounded but dial-able operating point (recall 0.165→0.763 at 1→10% false-positive rate), an elderly-cohort fairness tax and an asymmetric elder→young transfer collapse, a steep sensitivity to Doppler-axis scaling, calibration, determinism, and alarm fatigue — and deliver the killer cross-radar result: transfer across 24/77 GHz/UWB collapses to chance for both the token and the full model, so re-commissioning per radar is fundamental physics (Doppler scales with carrier frequency), not a token weakness. Two platform properties (one token composing classification and fall-novelty better than either subspace alone; per-capability energy falling as ~1/M) and five anchoring negatives — two per author prediction corrected by data — round out an honest, credible map. The token is not an accuracy champion and does not pretend to be; it is a predictable, auditable, private, self-healing cost instrument — exactly what the always-on layer on a privacy-preserving radar should be. This is a characterization of previously described, filed methods; it discloses no new algorithmic subject matter, and the per-deployment / per-radar selection of configuration is retained as trade secret. Highlights · The reframe — a Tier-0 layer on a privacy-preserving radar: the always-on radar tier (is this normal activity or a fall? is the room occupied?) is the contactless, camera-free instance of the sub-milliwatt layer beneath the NPU. The token decision is ~12 k operations and a measured 20 µs (49,000/s) atop a shared ~17 M-operation micro-Doppler front-end; int8-free; the raw radar cube never leaves the device. · A predictive accuracy map across two corpora / four radars: near-parity where few-class (Glasgow activity +0.022, coarse +0.006), a priced tax on fine confusable gaits (CI4R +0.06 to +0.08). And the near-parity is not leakage — it survives the strictest leave-one-collection-campaign-out split (+0.026, fully disjoint session/room/year/ people) and a permutation null collapses it to chance. · A label-free FALL gate (no labeled falls): mean AUROC 0.886, best-in-class at the safety-critical fall (per-class recall 0.959 ≥ the full model), and it beats every trivial "big/fast motion = fall" physics baseline (0.919 vs ≤0.81). · ⭐ Per-entity necessity REFUTED for falls: a population fall model (0.922) beats the per-subject self-twin (0.886) — a fall is a universal extreme event, so, unlike per-machine (#50) and per-patient (#48) anomalies, it needs no per-person commissioning. Per-entity necessity is conditional on the anomaly type — a new axis in the estate's map. · ⭐ The killer cross-radar result: transfer across 24 / 77 GHz / 10 GHz UWB collapses to chance for BOTH the token and a full model — a fundamental consequence of Doppler scaling with carrier frequency, not a token weakness → re-commission per radar (the same per-deployment posture as #44 cross-database, #48 per-patient, #50 per-microphone). · The honest deployment surface: a bounded-but-dial-able operating point (recall 0.165 / 0.546 / 0.763 @ 1 / 5 / 10% FPR); an elderly fairness tax (+0.062) and an elder→young transfer collapse (+0.154 — the token needs representative training); a steep Doppler-scale fragility (0.152 vs full 0.794 — a distance-codebook is scale-sensitive); calibration; bit-exact determinism; alarm fatigue (~321 wakes/day). · Two platform properties: the discriminant axis carries classification and the principal-component axis carries novelty, and composing them beats either axis alone on both (classify 0.834, novelty 0.907); per-capability energy falls as ~1/M. · Five anchoring negatives + two author predictions corrected by data (per-entity necessity; the false-alarm source is walking, not sit/bend). Honesty as evidence. · All characterization of filed / published methods — no new algorithmic subject matter; the boundaries are properties of the frozen method; the per-deployment / per-radar configuration-selection procedure is a trade secret. What this record contains · `Manuscript_Paper51.pdf` — the manuscript, ten figures embedded (accuracy/difficulty; the fall gate; the per-entity dissociation; deployment shift/fairness; the cross-radar collapse; fall-scrutiny vs detectors + physics baselines; the Doppler-scale corruption; per-class recall; the platform panel; the capability-composition + energy frontier), one table, and 73 references; and `Manuscript_Paper51.docx`, the editable source. · `PAPER_51_ZENODO_ARCHIVE.zip` — the reproducibility archive (md5 in ARCHIVE_MD5.txt): the frozen pre-registration, the runners (core; deployment-shift + cohort + population fall; the reviewer/acquirer stress battery; the skeptic battery; the capability-composition comparison; the CI4R cross-radar + difficulty + B1 analysis; the Modal fetch/processing apps; the CI4R fetch/featurize; the figure and POC builders), the frozen token encoder module, the per-experiment result records (JSON) and the CI4R key list, the ten figures, the manuscript source, and a README. All datasets are public and not redistributed (fetched at run time); all paths and identifiers are scrubbed (absolute paths → PATH_TO_DATA/PATH_TO_SCRATCH, any cloud handle → MODAL_USER) and leak-scanned. Cite as R. J. Ferlic and K. K. Ferlic, "The mmWave micro-Doppler Tier-0 gate: a deep, broad, and adversarially stress-tested map of a frozen single-token radar activity classifier and fall monitor," Zenodo, 2026, doi: [reserved at deposit]. License and patent notice Released under CC-BY 4.0. Consistent with that license, no patent or IP right of the authors is licensed, waived, or conveyed by this deposit. This work characterizes previously-described methods and discloses no new algorithmic subject matter; gradient boosting, k-means / vector quantization, Fisher discriminant analysis, micro-Doppler front-ends, temperature scaling, autoencoder / isolation-forest / one-class anomaly detection, 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 per-entity-necessity dissociation, the difficulty and fairness boundaries, and the Doppler-scale / cross-radar sensitivity are properties of the frozen filed method, not new subject matter.