The Configurable Bottleneck: A Cost-Constrained Capability Map for a Deterministic Edge Decision Token

Abstract

The Configurable Bottleneck: A Cost-Constrained Capability Map for a Deterministic Edge Decision Token Randolph James Ferlic, M.D. and Kimberly Kate Ferlic (Fieldstone Analytics, LLC, Austin, TX, USA) Preprint · Zenodo DOI: 10.5281/zenodo.22884023 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign Abstract A frozen, deterministic, class-discriminant single-token encoder (features → Fisher-discriminant ⊕ PCA subspace → k-means codebook of ≤ 256 cells → nearest-centroid decision) has been characterized piecemeal across a program of papers — its deployment behavior, its accuracy tax, its label-efficiency, its privacy — but never as one configurable platform, on one protocol, on the axes a deployment engineer actually trades. This is that unified, pre-registered, multi-axis, multi-domain characterization: a base token (M0) plus seven dials — a soft-vote readout (D1), a channel-partition token ladder (D2), a spiral (D3) and a diverse-model (D4) co-channel, a noise-robust reader (D5), a foundation codebook with shrinkage (D6), and a cheap quantized-sketch co-channel (D7), each a previously-filed or -published method — measured against the same base across ECG, industrial vibration, surgical-robot kinematics, wearable EMG, financial volatility, and continuous glucose data. The headline is deliberately honest and it reframes the object: on accuracy the token does not win — a cheap random-forest ensemble on the same features matches token-plus-co-channel, and a gradient-boosted model wins outright. What the token wins, by direct measurement, is cost: it reaches a decision in 3–5 microseconds per window versus 40–190 for that ensemble and 20–300 for the full model — 10 to 60× cheaper — as a fully enumerable ≤ 256-entry table that is auditable and never puts the raw signal on the wire. That silence on the wire is the token's largest cost advantage: against a full-data-stream-retention architecture it emits one byte instead of the raw window (an exact 180×–24,000× data reduction), and a pre-registered Monte-Carlo of radio-dominated device energy puts the transmission-energy savings at a median ≈ 350× (a Markov duty-cycle model brackets it from ≈ 14× vs an event-triggered device to ≈ 870× vs a continuous-streaming one; ≈ 900× even against the strongest decision-preserving codec) — one to two orders of magnitude beyond the compute figure. The platform is therefore a cost-versus-capability frontier: the bare token is the cheapest decision floor and concedes a real accuracy tax; the seven dials buy accuracy and robustness back, each at a stated compute, bit, or privacy cost, so a deployment spends only what its accuracy and threat budget require. Each dial's benefit generalizes across the six applications with a clear per-dial pattern (the diverse-model co-channel is the universal cheap accuracy/noise dial, +0.05–0.19; the foundation codebook is universal label-efficiency, near-ceiling at eight labels where a from-scratch codebook is at chance; the ladder recovers the multichannel accuracy tax, +0.12, and — its real headline — holds under sensor dropout where the bare token collapses; the noise-reader is additive-noise-specific). This generalization is confirmed by a pre-registered breadth battery on seven further physically-distinct modalities — inertial, chemical, acoustic/speech, gait, surgical-robotics, sleep/respiratory, and EEG — and at population scale (21,699 ECG records, tight CIs): the diverse-model co-channel is accuracy-positive on 8/8 and the foundation codebook label-efficient on 8/8, with one honestly-reported exception (a near-chance sleep/respiratory task where the dials cannot amplify absent signal). A clean-accuracy check further shows the expensive spiral co-channel recovers the accuracy tax only modestly and corruption-specifically, while the ladder's accuracy climbs with partition depth (one token per channel rivals/edges the diverse-model panel, at a stated bit cost) and — dually — a single token drives K parallel decisions at a negligible multi-task tax (mean +0.009 AUROC) and one-Kth the emitted rate (a sub-milliwatt core multiplexing several detectors for the price of one); and a cheap quantized-sketch co-channel matches that panel's accuracy while beating it on noise robustness at a fraction of the cost. The dials do not simply stack: composition is threat-dependent, one tempting combination (ladder + spiral) actively conflicts under compound stress, and — pre-registered and reproduced on a second raw multichannel domain (PTB-XL 12-lead) — the best configuration is threat-specific — and running all dials together is the worst point on the frontier (a spiral-dominated sum of costs, 1.4–3.3 ms, for at most a +0.025 bump that plateaus after the second co-channel), the rule being to compose across orthogonal axes, never stack within one. The token is robust to the corruptions a low-power sensor actually imposes (powerline interference, int8 quantization) at zero added cost; identity leakage from the emitted token falls toward chance as the population grows (an entropy cap), and a decision co-channel adds no identity while an identity-rich raw one does; and — honestly — when compute is unconstrained a single fixed dial can equal the best-per-threat oracle, so the map's value is precisely cost-constrained: the cheapest dial that meets each requirement. Findings are stable across the readout, codebook-size, and noise settings. The deployment recipe that follows is the platform's actual value proposition: a single deterministic, auditable, sub-milliwatt core that a partner positions anywhere on the cost↔capability frontier — per site, per threat, per power budget. This is a characterization of previously-described, filed methods; it discloses no new algorithmic subject matter, and the per-deployment selection of dials and their settings is retained as trade secret. Highlights · The honest reframe: the token is not the accuracy story — it is the cost story. Answering the bluntest reviewer attack ("did you just add a random forest?") with measurement: a cheap diverse-model panel on the same features matches token+co-channel and a gradient-boosted model wins outright. The token's moat is measured cost — 3–5 µs/window versus 40–190 (panel) and 20–300 (full model), 10–60× cheaper — as an enumerable ≤ 256-entry auditable table that keeps the raw signal off the wire. · One core, seven dials, one cost↔capability frontier. The bare token is the cheapest decision floor and concedes a real accuracy tax; seven previously-filed/-published dials buy accuracy, drift- and noise-robustness, fault-tolerance, and label-efficiency back, each at a stated compute / bit / privacy cost. A deployment spends only what its budget requires. · Each dial's trade-off generalizes — with a per-dial pattern. Across six domains: the diverse-model co-channel (D4) is the universal cheap accuracy/noise workhorse (+0.05–0.19 at ~16 µs, ~65% of the spiral's benefit at ~1/60th its compute); the foundation codebook (D6) is universal label-efficiency (population ceiling at eight labels where a fresh codebook is at chance, +0.33–0.48); the soft-vote (D1) helps every non-saturated domain for free; the ladder (D2) recovers the multichannel accuracy tax (+0.12); the noise-reader (D5) is additive-noise-specific. · The ladder's real headline is fault-tolerance. On a saturated single-condition bench the ladder shows nothing — but under sensor dropout the bare token collapses (macro-AUC 0.70 on CWRU, 0.56 on PTB-XL) while the channel-ladder holds (1.00 and +0.062): the distributed multi-token code survives channel loss where the single token cannot. Reproduced on both raw multichannel domains. · The dials do not stack — the best choice is threat-specific. A pre-registered 2×2 of drift × sensor-dropout, reproduced on CWRU and PTB-XL 12-lead: clean → bare token; drift → free soft-vote; dropout → fault-tolerant ladder; compound drift+dropout → a single token + spiral co-channel — which beats ladder+spiral (the composition conflict: the ladder's drift-hit surviving-channel token corrupts the equal-weight fusion). Select one dial by the threat; do not stack. · Robust to the corruptions a sub-milliwatt sensor actually sees, and privacy in-frame. Inherent robustness to powerline interference (60 Hz) and int8 quantization at zero added cost and no dial needed; broadband noise / baseline wander / missing samples are rescued by a co-channel or soft-vote. Emitted-token identity leakage falls toward chance as the population grows (an entropy cap), and a decision co-channel adds no identity while an identity-rich raw one does. · Honest about the map's own value. When compute is unconstrained a single fixed dial can equal the best-per-threat oracle — so the map is valuable precisely under a budget: it names the cheapest dial that clears each requirement. Findings are stable across readout (m), codebook-size (K), and noise (σ) settings. What this record contains · Manuscript_Paper47.pdf — the manuscript with eight figures embedded, in reading order (Figure 1 the cost↔capability frontier, Figure 2 the measured-cost bars, Figure 3 the data-movement/energy-savings distribution, Figure 4 the cross-domain generalization heatmap, Figure 5 the cross-modality breadth chart, Figure 6 the ladder partition-depth curve, Figure 7 the two-domain composition map, and Figure 8 the consolidated D.0–D7 scorecard), and Manuscript_Paper47.docx, the editable source. · PAPER_47_ZENODO_ARCHIVE.zip — the reproducibility archive (md5 in ARCHIVE_MD5.txt): the eleven frozen pre-registrations (the dial-map protocol, the R0–R7 reviewer-proofing battery, two Tier-2 breadth pre-registrations, and seven refinement pre-registrations — spiral clean-accuracy, ladder partition-depth/sketch, multi-task/parallel-decisions, scorecard-completion, dial-combinations, and the Monte-Carlo and Markov energy models), the dial-map / reviewer-pr