Decision-Bits, Not Fidelity-Bits: A Rate–Accuracy Comparison of a Task-Aware Single-Token Encoder Against General-Purpose Compression Codecs

Decision-Bits, Not Fidelity-Bits: A Rate–Accuracy Comparison of a Task-Aware Single-Token Encoder Against General-Purpose Compression Codecs Randolph James Ferlic, M.D., and Kimberly Kate Ferlic — Fieldstone Analytics, LLC Correspondence: [email protected] Preprint · Zenodo DOI: 10.5281/zenodo.21982600 · CC-BY 4.0 Abstract General-purpose compression codecs such as ZFP and SZ3 are reconstruction codecs: they spend bits to recover the input waveform to a bounded fidelity, and are judged by rate–distortion. Many edge and on-device applications, however, need only a decision from each signal window (a fault class, an arrhythmia flag), not the waveform itself. For such tasks the relevant question is not how faithfully a signal is reconstructed but how well a decision is preserved per bit — a rate–accuracy criterion. We compare a deterministic, task-aware single-token encoder against four general-purpose codecs (ZFP, SZ3, a DCT top-k transform codec, and lossless zstd) and a learned autoencoder codec on this criterion. Across four public benchmarks in three modalities — industrial bearing vibration (CWRU 10-class within-load and cross-load; MFPT 3-class), ECG abnormal-beat screening (MIT-BIH), and subject-disjoint da Vinci surgical-robot kinematics (ROSMA 3-class) — the task-aware token preserves the decision at about 8 bits/window, within roughly 0.01–0.03 AUC of the uncompressed ceiling on the bearing and ECG tasks (a larger ~0.11 on the hardest, high-dimensional surgical task), while the best general-purpose codec requires 224× to 15,656× more bits to reach the same decision accuracy. The dominance survives a pre-registered battery of controls: retraining the codec's classifier on its own reconstructions (its best case) still leaves the token 69–605× ahead; the result holds across three classifier heads, across codebook sizes, for an unsupervised token, and against a learned neural codec (which fails to reach the token within the swept bitrate on the bearing and ECG tasks, and closes the gap on the hardest surgical task only until a few more tokens are spent — four channel-partition tokens climb to 0.88 AUC and surpass the learned codec's best point at 6.4× fewer bits). We situate the finding in the theory of minimal sufficient statistics and the goal-oriented (coding-for-machines) communication literature, and release a fully reproducible protocol. All results use real public data; the comparison favors the general-purpose codecs where any ambiguity exists, so the reported dominance is conservative. Highlights • When the receiver is a machine that only needs a decision, the right axis is rate–accuracy (decision AUC per bit), not rate–distortion: a deterministic task-aware single token preserves the decision at about 8 bits per window. • Against four general-purpose codecs (ZFP, SZ3, a DCT top-k transform codec, lossless zstd) and a learned autoencoder, the best codec needs 224× to 15,656× more bits to reach the same decision accuracy — two to four orders of magnitude. • The dominance survives every pre-registered fairness control, including retraining the codec's own classifier on its reconstructions (still 69–605× ahead), three classifier heads, an unsupervised token, and a codebook-size sweep. • The single-token accuracy tax grows with task information (about 0.01 AUC on bearing vibration, 0.03 on ECG, 0.11 on the hardest subject-disjoint surgical task) — a clean rate–relevance datapoint; dominance and the tax are separate axes. • Token count is a tunable operating point: on the hardest task a few channel-partition tokens (32 bits) climb to near-ceiling accuracy and surpass the learned codec's best point at 6.4× fewer bits. What this record contains • The manuscript (PDF). • A reproducibility archive (`PAPER_36_ZENODO_ARCHIVE.zip`): deterministic pre-registered runners, the per-run JSON result summaries behind every figure and table, the four schematic/data figures, and a frozen pre-registration protocol with a per-study run-log. All datasets are public (CWRU and MFPT bearing vibration; MIT-BIH ECG via PhysioNet; ROSMA da Vinci surgical kinematics, Zenodo 10.5281/zenodo.3932964); SZ3 is compiled from source. No raw subject data is redistributed. Cite as R. J. Ferlic and K. K. Ferlic, "Decision-bits, not fidelity-bits: a rate–accuracy comparison of a task-aware single-token encoder against general-purpose compression codecs," Zenodo, 2026, doi:10.5281/zenodo.21982600. License and patent notice Released under CC-BY 4.0. Consistent with Section 2(b) of that license, no patent, patent-application, or other intellectual-property right of the authors is licensed, waived, granted, or otherwise conveyed by this publication or by any reuse of it; the encoding methods described herein — including the single-token and the multiple-token (product-quantization / channel-partition) embodiments — are the subject of pending U.S. patent applications. Companion deposits Single-token class-discriminant codebook family on Zenodo (community: spiral-domain-encoder-campaign), including the encoder-method deposit and the on-device drift-adaptation studies. Keywords task-aware compression, rate–distortion, rate–accuracy, coding for machines, goal-oriented communication, edge inference, sufficient statistic, information bottleneck, SZ3, ZFP, autoencoder, product quantization, condition monitoring, ECG, surgical robotics, da Vinci Research Kit, pre-registration.