The Price of the Bottleneck: A Pre-Registered, Multi-Domain Deployment Characterization of a Deterministic Edge Decision Token

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

The Price of the Bottleneck: A Pre-Registered, Multi-Domain Deployment Characterization of 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.22838118 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign Abstract Compact, deterministic "decision tokens" — a window of a sensor stream quantized on-device to a single ~8-bit code that carries the byte a downstream decision needs — are attractive for edge and ambient intelligence: bandwidth and energy collapse, the raw signal never leaves the device, and the decision is auditable. Accuracy for such encoders is routinely reported; their deployment behavior — will it hold up on a different device, site, population, or over time; when should it be trusted; is it fair; does it survive corrupted signal; how much data does it need; is it reproducible — is not. We give a pre-registered, adversarial, multi-domain deployment characterization of one concrete instance: a frozen, deterministic, class-discriminant single-token encoder (features → Fisher-discriminant ⊕ PCA subspace → k-means codebook of ≤ 256 cells → nearest-centroid decision). We test six deployment axes — drift, calibration / out-of-distribution (OOD) / abstention, subgroup fairness, corruption robustness, sample efficiency, reproducibility — each with frozen Go/No-Go reads, stress the findings through four hardening passes (strong baselines, adversarial evasion, statistical confidence intervals, honest re-evaluation), and replicate the core axes across five application domains (healthcare ECG, industrial bearing vibration, surgical-robot kinematics, wearable EMG, financial volatility). The result is a sharp, honest map. The token is bit-exact reproducible and its decision function is a ≤ 256-entry auditable table; it is stable across acquisition device, site, era, and patient demographics, discriminatively fair across sex/age/device, and in-distribution calibrated. Its operating envelope is equally specific and, with confidence intervals, statistically characterized: it drifts significantly worse than a modern model only on unseen individuals in identity-entangled tasks (ECG); it is more label-hungry with no few-shot/transfer shortcut; it is brittle to heavy noise and missing input channels; it cannot recognize a novel class it was not trained on (universal across every free confidence signal); and "gate on novelty" is not a deployable general OOD gate — it filters gross corruption only. We trace the cross-individual drift to its mechanism: it is not a removable identity representation (three principled front-ends — identity-axis projection, adversarial patient-invariant encoding, cross-individual augmentation — all fail to fix it) but a distributional limit of the discrete codebook, and it is largely recoverable post-hoc by a soft, distance-weighted readout over the nearest cells (inter-patient drift falls ~72%, label-free, token unchanged). Finally, an integrated deployment simulation shows the recommended confidence-gated pipeline cuts the confident-wrong rate ~4–6× but abstains on ~60% of inputs and still mis-decides genuinely novel classes — so the honest deployment posture is a high-abstention, human-in-the-loop assistant, not an autonomous decider. Results that changed under adversarial hardening are reported as corrections. This is explicitly a characterization of previously-described, filed methods; it discloses no new algorithmic subject matter. Highlights · A deployment map, not a benchmark score — six pre-registered deployment axes (drift, calibration/OOD/abstention, fairness, corruption, efficiency, reproducibility), each with a frozen Go/No-Go read, plus four adversarial + statistical hardening tiers that deliberately attack the paper before it is written. What is measured is deployment behavior, the thing a partner needs before a pilot, not accuracy alone. · The strengths, believable because the limits are named — the frozen token is bit-exact reproducible (float32), its whole decision function is a ≤ 256-entry auditable table, it is stable across device / site / era / demographics, discriminatively fair (AUROC gap ≤ 0.05 on sex/age/device), and in-distribution calibrated (ECE ≈ a logistic regression). · The honest limits, sized with confidence intervals — more label-hungry (~512 vs ~32 labels on a hard task) with no transfer shortcut; brittle to heavy noise and missing channels; blind to novel classes (novelty AUROC 0.57, universal across all four free signals); and a novelty gate that catches gross corruption only (90% OOD recall costs ~70% of clean inputs). · The identity-entanglement through-line — cross-individual drift is significant (worse than a modern model) only on ECG, where the decision is entangled with patient morphology (token−RF drift CI [0.016, 0.325], excluding 0); on robotics, wearable, and financial tasks the token is statistically indistinguishable from a modern model on drift. The biggest limitation bites precisely and only when the decision is entangled with who the individual is. · A modest cross-database tax + a preprocessing-mismatch sensitivity (corrected). A wide four-site matrix (PTB-XL/DE, CPSC/CN, Georgia/US, Chapman/CN, uniform preprocessing) shows the codebook transfers across sites with no collapse — cross-site AUROC median 0.735, ~0.05 worse than a modern model. A controlled test (same records, only sampling-rate handling varied) isolates a token-specific preprocessing-mismatch sensitivity (0.044 drop vs the modern model's ~0). We retract, verbatim, an earlier two-site “collapse to 0.539” that the controlled matrix showed to be an unrepresentative outlier. Rule: harmonize/pin the preprocessing pipeline across sites and budget a small per-site recalibration — a manageable tax, not a barrier. · Co-channels: a second post-bottleneck mitigation, at a privacy cost. A small co-channel ensemble (K parallel tokens from feature partitions — the filed channel-partition ladder, averaged post-bottleneck) cuts MIT-BIH inter-patient drift ~half (+0.116→+0.057 at K=2) and recovers int8 on-device decision agreement 0.65→0.98; but emitting those co-channel tokens sharply multiplies re-identifiability (companion privacy paper) — the same channel-partition knob is a deployment lever and a privacy risk, so compute co-channels on-device and egress only the decision. · The price of the bottleneck — and its refund — the cross-individual drift is not a removable identity representation (identity-axis projection cleanly de-identifies at zero decision cost but does not reduce drift; domain-adversarial invariance and cross-individual augmentation also fail) but a distributional property of the discrete codebook. It is largely recoverable post-hoc by a distance-weighted soft readout over the m nearest cells — inter-patient drift falls ~72%, with no target labels and the emitted 8-bit token unchanged — and m is a tunable knob on the purity ↔ generalization tradeoff. · The integrated deployment number — running the full recommended pipeline (novelty gate → margin abstention → decide) on a mixed stream cuts the confident-wrong rate ~4–6× (0.32 → 0.07), but abstains on ~60% of inputs and still lets genuinely novel classes through. The safe posture is a high-abstention, human-in-the-loop assistant. · The honest bound, held deliberately conservative — the hardening tiers set the envelope: the token is not claimed to beat modern models on cross-device transfer (not significant), the novelty gate is scoped to gross corruption, the fairness gap is sized out-of-sample (0.23 → 0.19, not 0.03), the int8 reproducibility claim is qualified, and MIT-BIH is treated as the single-feature task it is. What this record contains · Manuscript_Paper44.pdf — the manuscript with four figures embedded (the deployment scorecard, the post-bottleneck drift fix, the integrated deployment tradeoff, and the wide cross-database matrix + preprocessing-isolation), and Manuscript_Paper44.docx, the editable source. · PAPER_44_ZENODO_ARCHIVE.zip — the reproducibility archive: the frozen pre-registrations, the experiment runners (the six gap studies, the four hardening tiers + deploy stress test, the five-domain replication + bootstrap CIs, and the drift-mechanism probes — identity-suppression, domain-adversarial, scramble, and post-bottleneck soft readout), the cross-database transfer study (PTB-XL ↔ Chapman, streamed from PhysioNet), the per-experiment result records, the shared encoder loader, the consolidated summary scorecard, and a README. All datasets are public; no raw benchmark data is redistributed (sources and a `PATH_TO_DATA` convention are in the README). All paths/identifiers are scrubbed and leak-scanned per the campaign deposit discipline. Cite as R. J. Ferlic and K. K. Ferlic, "The price of the bottleneck: a pre-registered, multi-domain deployment characterization of a deterministic edge decision token," Zenodo, 2026, doi: 10.5281/zenodo.22838118. License and patent notice Released under the Creative Commons Attribution 4.0 International License (CC-BY 4.0). Consistent with that license, no patent, patent application, or other intellectual-property right of the authors is licensed, waived, granted, or otherwise conveyed by this deposit. This work characterizes the deployment behavior of previously-described methods and discloses no new algorithmic subject matter. The methods characterized — the class-discriminant single-token codebook encoder and its nearest-centroid monitor, the identity-axis-projection front-end evaluated in the drift-mechanism study, and the multi-token / soft-readout mechanisms referenced — 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 multi-token / token-ladder readout ap