Three Composable Sensor-Encoder Architectures Across the Decision-Fidelity Spectrum: The Class-Discriminant Codebook, the Logarithmic Spiral-Domain Transform, and Their Selective Dual-Path

Across a multi-year campaign we developed and validated three distinct sensor-encoder architectures, each occupying a different point on a decision–fidelity spectrum. The class-discriminant codebook compresses each window of a signal to a single decision-oriented token — extreme compression that preserves the *decision* while discarding the waveform — and supports, from one token stream, classification near an uncompressed ceiling, anomaly detection, multi-task inference, entropy coding, and privacy by construction. The logarithmic spiral-domain transform is a closed-form, deterministic, bijective re-coordinatization that recovers the *exact waveform* to machine precision (mean absolute error of order 1e-9 to 1e-18 on real signals) but, by itself, does not compress. The selective tiered-fidelity dual-path composes the two in parallel: the codebook delivers the decision for every window, while the spiral co-channel is invoked on demand — routed by the codebook's own anomaly monitor — to recover the exact waveform only for the windows that warrant it. This paper places the three on a common axis (with a spectrum diagram, an architecture schematic, and a qualitative attribute radar), contrasts what each optimizes, synthesizes the campaign's headline validated results, gives design guidance for choosing among them, and states the attribution honestly: compression and privacy are properties of the codebook, lossless invertibility is a property of the spiral, and the selective dual-path is the architecture that lets a single deployment have both — the decision everywhere and the exact waveform exactly where a monitor says it matters. The campaign's honest negatives are restated so the boundaries of each architecture are explicit. No new experiments are reported; the contribution is the unification, the design guidance, and the corrected attribution. Keywords / index terms: sensor encoding; class-discriminant codebook; logarithmic spiral-domain transform; selective reconstruction; tiered fidelity; dual-path encoder; decision–fidelity tradeoff; lossless reconstruction; edge sensing; encoder taxonomy; honest attribution. Selected references: information bottleneck (Tishby et al. 1999); vector quantization (Linde-Buzo-Gray 1980); task-oriented/semantic communication (Gündüz et al. 2023; Shao et al. 2022); scalable video coding (Schwarz et al. 2007); JPEG2000 (Taubman & Marcellin 2002); neural image compression (Ballé et al. 2017); selective classification (Geifman & El-Yaniv 2017); out-of-distribution detection (Hendrycks & Gimpel 2017); and the companion deposits Papers 1, 19–27. Companion deposits in this Zenodo Community (spiral-domain-encoder-campaign): · Paper 19 — 10.5281/zenodo.20788187 · Paper 20 — 10.5281/zenodo.20802759 · Paper 21 — 10.5281/zenodo.20802826 · Paper 22 — 10.5281/zenodo.20805321 · Paper 23 — 10.5281/zenodo.20821668 · Paper 24 — 10.5281/zenodo.20821779 · Paper 25 — 10.5281/zenodo.20821903 · Paper 26 — 10.5281/zenodo.20854722 · Paper 27 — 10.5281/zenodo.20869093