Where the Cheapest Token Wins, Pays, and Fails: A Predictive, Stress-Tested Map of a Frozen Single-Token Edge Encoder

Where the Cheapest Token Wins, Pays, and Fails: A Predictive, Stress-Tested Map of a Frozen Single-Token Edge Encoder Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: 10.5281/zenodo.22945419 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign Abstract The 2026 wave of edge artificial intelligence optimizes the expensive tier: new mobile silicon runs generative, agentic models on-device at unprecedented performance-per-watt. Almost nothing has been done to make the always-on sensing-and-decision tier — the layer that decides when the expensive tier should even wake — comparably cheap. We study a frozen, deterministic, class-discriminant single-token encoder (features → 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) as a candidate Tier-0 layer that sits beneath the neural processing unit (NPU). One decision costs on the order of 20,000 operations — roughly four to five orders of magnitude fewer than a single generative token — and runs on a sensor hub, digital signal processor, or microcontroller with no NPU; priced in published per-operation energies, its arithmetic is tens of nanojoules or less per decision, roughly two orders of magnitude leaner than a deployed keyword-spotting wake-up. Rather than argue merely that such an encoder "works," we build and adversarially stress a predictive map of where it wins, pays a cost, and fails. Across ten physically distinct public datasets and a same-frozen pipeline, two axes emerge. On the accuracy axis, the token is competitive-to-winning against a strong gradient-boosted model where discriminative information lives in per-channel time-frequency statistics (bearing vibration: +0.16 to +0.21 AUROC; surgical kinematics: +0.067; electrocardiography: within +0.006), pays a small-to-moderate cost tax where the signal is per-channel but harder (surface electromyography ≈ 0.05–0.10; financial volatility-regime detection ≈ 0.12), and fails where the discriminative information is fundamentally spatial (electroencephalographic motor imagery: near chance, features discard the spatial code). On the personalization axis, a label-free per-entity "self-twin" recalibration recovers drift that is a local-distribution shift (sensor re-donning, cross-subject, cross-machine-load, cross-surgeon: full or 70–88% recovery) but fails where the non-stationarity is stationary-but-noisy (financial regime "drift," where more history beats recent data). We subject every load-bearing claim to four pre-registered adversarial rounds (twenty-plus attacks), fold each concession, and anchor the map with two genuine negatives (spatial-signal failure; self-twin failure). We further reconcile an apparent contradiction — that electrocardiography is simultaneously a "cheapest digital twin" showcase and a "cost-tax" domain — by showing the two claims measure different axes and are consistent. The token is, and remains, a cost and label-efficiency instrument, not an accuracy champion; its value on the agentic edge is a sub-milliwatt, auditable, private, self-healing Tier-0 layer whose behavior is now predictable from a domain's signal structure. This is a characterization of previously described, filed methods; it discloses no new algorithmic subject matter, and the per-deployment selection of configuration is retained as trade secret. Highlights · The reframe — a Tier-0 layer beneath the NPU: the 2026 silicon makes the expensive generative tier cheap; this positions the frozen token as the always-on tier below it — one decision ≈ 20,000 operations, roughly 10⁴–10⁵× fewer than a single generative token (tens of nanojoules or less per decision when priced in published per-operation energies, ~2 orders of magnitude leaner than a deployed keyword-spotting wake-up), on a sensor hub with no NPU, one byte on the wire. · The deliverable is a PREDICTIVE map, not a demo: a domain's signal structure predicts the token's behavior. It wins where information is per-channel time-frequency (bearing vibration +0.16/+0.21, surgical kinematics +0.067), is near-parity on electrocardiography (+0.006, the smallest tax), pays a bounded tax where per-channel but harder (sEMG ≈ 0.05–0.10, financial 0.117), and fails where the signal is fundamentally spatial (EEG motor imagery ≈ chance). · A label-free self-twin recovers LOCAL-distribution drift: cross-day surface electromyography (70–88%), cross-load bearing (full), cross-surgeon surgical (full), and per-patient electrocardiography (94% of a supervised ceiling without labels) — but fails on stationary-but-noisy financial regime non-stationarity, where more history beats recency. · Two genuine negatives anchor the map: a spatial-signal failure (EEG motor imagery, where a strong model on the same per-channel features fails too — the feature front-end, not the token) and a self-twin failure (financial). Honesty as evidence: a characterization that never reports a failure has not been stressed. · The target clinical population, tested: on eleven trans-radial amputees the encoder decodes intent (~4–6× chance), de-risking technology fit — while honestly bounding it (≈2.2× harder, a doubled tax, a useless cross-amputee population codebook, and label-hungry per-individual personalization). · Free-dial stack + multiplexing: an edge-default M0+D6+D1 stack (runtime-free/near-free, best label-efficiency, a built-in gate confidence), and one sub-milliwatt core driving up to K=40 heterogeneous detectors at ≈ 0 accuracy tax, with per-decision compute and emitted bits both falling as ≈1/K. · Four pre-registered adversarial rounds (20+ attacks): every load-bearing claim defended or conceded verbatim, including a methodology-leakage audit that exonerates the headline (rep/subject-disjoint) results; the ECG two-axis reconciliation in a single table. · All characterization of filed / published methods — no new algorithmic subject matter; the per-deployment configuration-selection procedure is a trade secret. What this record contains · `Manuscript_Paper49.pdf` — the manuscript, seven figures embedded (the two-tier architecture; the per-dial edge cost tiers; the K-detector multiplexing economics; the predictive two-axis map; the self-twin recovery across modalities; the token-versus-full-model bars; the electrocardiography two-axis), eight tables, and 73 references; and `Manuscript_Paper49.docx`, the editable source. · `PAPER_49_ZENODO_ARCHIVE.zip` — the reproducibility archive (md5 in ARCHIVE_MD5.txt): the frozen pre-registrations (the free-dial stack, the four data-in-hand adversarial rounds, the five new-dataset stresses, the ECG two-axis, and the same-day / multi-day self-twin pre-regs), the runners (the Tier-0 op-count and dial/multiplexing economics, the free-dial- stack validators, the four adversarial-round runners, the new-dataset stress runners, the ECG two-axis runner, the two self-twin runners, and the figure builders), the frozen token encoder module, the per-experiment result records (JSON) and per-simulation result notes (markdown), the seven figures, the manuscript source, and a README. All datasets are public and not redistributed (fetched from their public sources 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, "Where the cheapest token wins, pays, and fails: a predictive, stress-tested map of a frozen single-token edge encoder," Zenodo, 2026, doi: 10.5281/zenodo.22945419. 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, product quantization, Fisher discriminant analysis, empirical-Bayes shrinkage, temperature scaling, 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 / ladder and soft readouts, inference-time 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 / ladder application (No. 64/119,487), and the inference-time fusion application (No. 64/137,805). The per-deployment selection of configuration is retained as a trade secret and is not disclosed here. © 2026 Fieldstone Analytics, LLC and the authors. Inquiries: [email protected]. Companion deposits (spiral-domain-encoder-campaign) · The Configurable Bottleneck (the platform whose dials are edge-characterized here): doi:10.5281/zenodo.22884023 · The Cheapest Digital Twin (the self-twin whose scope this maps): doi:10.5281/zenodo.22922678 · Class-discriminant codebook construction (the base encoder): doi:10.5281/zenodo.20788187 · Deterministic multi-token token ladder: doi:10.5281/zenodo.22003179 · Label-free inference-time channel fusion: doi:10.5281/zenodo.22046713 · 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 price of the bottleneck (deployment): doi:10.5281/zenodo.22838118 · Paying down the price of the bottleneck: doi:10.5281/zenodo.22866