A Multi-Task Substrate: Serving Plural Downstream Tasks from a Single Shared Class-Discriminant Codebook Token Stream

A device that compresses a continuous signal to a class-discriminant codebook token stream often must make several decisions on the same signal — multiclass classification, binary screening, per-condition detection, regression, anomaly monitoring. Instantiating a separate encoder-and-model pipeline per task multiplies the encoder computation, memory, and energy that dominate an edge footprint. We show a multi-task substrate: a single shared codebook and its token stream are read by a plurality of concurrent decoupled heads, each performing a distinct task within a small tolerance of a task-specific codebook, so that an added task costs only a lightweight head and no additional encoding. On real twelve-lead electrocardiography one shared codebook served five-class diagnosis, binary screening, and age regression each within tolerance of a dedicated codebook (gaps ≈ 0.000 / 0.017 / 0.028), and on inertial human-activity data at parity. The per-window encoder cost was ≈ 13.8 µs versus ≈ 0.015 µs for a head — ≈ 0.1% — so ten tasks from one shared encoder ran ≈ 9.9× faster than ten separate encoders, and an added task was effectively free. The substrate is head-architecture-agnostic (lookup, logistic, neural, tree within ≈ 0.004). A task whose discriminative structure is orthogonal to a single fitting objective (e.g., a demographic attribute relative to diagnosis) is *not* served by a diagnosis-only codebook; fitting the shared codebook on a composite multi-task objective (a union of the target tasks) recovers such a task (gap ≈ 0.093 → 0.046, a statistically significant recovery) while preserving the aligned tasks. The operative lever is the fitting objective, not subspace size, and growing the composite union to five tasks did not degrade the included tasks. We report honest bounds verbatim: a task excluded from the composite fit and orthogonal to it is not served zero-shot; per-patient personalization increases detection sensitivity but at the cost of additional false alarms (detection AUC ≈ 0.851 → 0.942 but false-alarm ≈ 0.140 → 0.777 — a disclosed tradeoff); and serving a demographic task from the shared codebook leaks modestly more of that attribute (≈ 0.608 → 0.656, below a material threshold, since diagnosis already correlates with demographics). The two cost regimes are characterized: an in-fit task adds only a head (≈ 130 µs, free) while a new orthogonal task incurs a codebook re-fit (≈ 463 ms). The contribution is the one-stream multi-task substrate, the composite-fitting mechanism that broadens it, and the measured limits of both — a read-only addition to an existing class-discriminant token pipeline. Keywords / index terms: multi-task learning; shared representation; class-discriminant codebook; token stream; compute efficiency; composite fitting; negative transfer; regression; cross-modality; per-patient personalization; demographic privacy; pre-registration; honest negatives. References: 1. R. Caruana, "Multitask learning," Machine Learning, 1997. 2. S. Ruder, "An overview of multi-task learning in deep neural networks," arXiv:1706.05098, 2017. 3. Y. Zhang and Q. Yang, "A survey on multi-task learning," IEEE TKDE, 2021. 4. T. Standley et al., "Which tasks should be learned together in multi-task learning?," ICML, 2020. 5. R. A. Fisher, "The use of multiple measurements in taxonomic problems," Annals of Eugenics, 1936. 6. P. Wagner et al., "PTB-XL, a large publicly available electrocardiography dataset," Scientific Data, 2020. 7. G. Moody and R. Mark, "The impact of the MIT-BIH arrhythmia database," IEEE EMB Magazine, 2001. 8. D. Anguita et al., "A public domain dataset for human activity recognition using smartphones," ESANN, 2013. 9. R. J. Ferlic and K. K. Ferlic, companion deposits (Papers 19, 20, 21, 22, 23, 24), Zenodo, 2026. 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