The Composed Platform: one pre-deployment screen predicts a multi-capability frozen-token stack under cross-domain stress — with an axis-specific reach, on commodity silicon

Abstract

The Composed Platform: one pre-deployment screen predicts a multi-capability frozen-token stack under cross-domain stress — with an axis-specific reach, on commodity silicon Randolph James Ferlic, M.D. and Kimberly Kate Ferlic — Fieldstone Analytics, LLC, Austin, TX, USA Preprint · Zenodo DOI: 10.5281/zenodo.23401467 · CC-BY 4.0 · Community: spiral-domain-encoder-campaign · a characterization of previously filed and published methods; no new algorithmic subject matter is disclosed. Abstract A frozen, deterministic single-token encoder exposes nine composable capabilities — accuracy, robustness, and monitoring — on one runtime, and a prior program has shown that a single class-separability measurement, computed once on a deployment's commissioning data, predicts the encoder's single-capability behavior across domains. This paper answers the question a technical evaluator actually asks: does the composed stack hold under cross-domain stress, and can one cheap probe de-risk the whole platform at once? We run a pre-registered study across sixteen domains spanning seven modality families and four stressors, with five falsifiable hypotheses and frozen bands. The central finding is an honest boundary, not a blanket win. Composition is free and deterministic. Across ninety-six capability-additions under stress, 100% are non-destructive (0 destructive beyond a max(0.01, seed-spread) margin), and the composed stack is bit-exact — 0 token flips on the host, anchored to 0/256 flips on two microcontroller architectures (float32 and int8). One screen predicts the composed safety gate — and its reach is axis-specific. A single commissioning-time separability measurement predicts the composed safety gate (Spearman +0.92, permutation p = 0.0002, thirteen domains), and that reach is axis-specific: the same screen is blind to corruption-robustness (Spearman −0.23, p = 0.39 — refuted exactly as pre-registered). The screen answers the accuracy, label-budget, and safety-gate questions, because those are properties of where the classes sit; it does not answer the corruption-robustness question, which is a property of how the input is degraded and needs its own input-conditioning dial. The honest mitigation limit. Label-free recentering recovers only ~30% of cross-domain loss (not a majority), so the deployment consequence is a requirement, not a defect: under drift the stack must be recalibrated per-epoch, and label-free recentering is a partial mitigation between recalibrations. It runs on commodity silicon, measured and hardened. On an Arm Cortex-M33 the frozen runtime is bit-exact (0/256 flips across two architectures, two models, and both numeric paths, replicated across twelve power-on resets); its worst-case inference latency is 1.573 ms with near-zero jitter (a spike, not a spread, 3× inside a 5 ms budget); a continuous 50,678-inference soak produced zero token changes and zero faults; per-inference energy is 9.73 ± 0.07 µJ on a sub-milliwatt idle floor; and on-device compute costs less than a single BLE advertising event (16.6 µJ, including the CPU wake and three-channel transmission), so shipping the one-byte token instead of the raw window saves roughly 20–70× in radio energy. Highlights · A composable platform, not features that fight each other: 96/96 capability-additions non-destructive under cross-domain stress; 0 token flips (0/256 on silicon). · One cheap commissioning probe predicts the composed safety gate (Spearman +0.92, p = 0.0002) — the program's separability law survives composition. · The honest boundary that makes the probe deployable: the same screen is blind to corruption-robustness (Spearman −0.23). Run the probe for the class-geometry axes; condition the input for the rest. · A stated requirement, not a hidden weakness: label-free recentering recovers only ~30% of cross-domain loss → recalibrate per-epoch under drift. · A measured, hardened silicon envelope: WCET 1.573 ms (≈0 jitter), a 50,678-inference zero-drift soak, 9.73 ± 0.07 µJ/inference, compute cheaper than one radio packet, 12.7 KB footprint, sub-milliwatt. · The economics of one byte (the deployment implications): on-device compute moves the energy bottleneck off the radio — ~25× lower energy per delivered decision and ~150–4,000× smaller transmit/store footprint, on a ~$10-class commodity microcontroller, with the raw waveform never leaving the device. · One platform, nine composable capabilities, five invariants: a deployment composes a subset on one runtime; Tier-0 on-device execution is the fifth architectural invariant (proven on silicon), and AI-safety runtime monitoring is the flagship application of the self-twin capability — not separate capabilities. · Two of five pre-registered hypotheses refuted — each refutation sharpening the claim rather than undermining it. What this record contains · Manuscript_Paper62.pdf / .docx — the manuscript (7 figures, 30 references). · PAPER_62_ZENODO_ARCHIVE.zip — the reproducibility archive: the frozen pre-registration (five hypotheses with a-priori bands), the composition-under-stress runner, the shared dataset loaders + the frozen token encoder + the composed-stack battery, the per-domain result records (JSON) and aggregate verdicts, the seven figures and their builder, and a README. Public corpora are not redistributed; absolute paths, handles, and productization know-how are scrubbed. Cite as Ferlic, R. J., & Ferlic, K. K. (2026). The Composed Platform: one pre-deployment screen predicts a multi-capability frozen-token stack under cross-domain stress — with an axis-specific reach, on commodity silicon. Zenodo. https://doi.org/10.5281/zenodo.23401467 License and patent notice Released under CC-BY 4.0. Consistent with that license, no patent or other intellectual-property right of the authors is licensed, waived, or conveyed by this deposit. This work characterizes previously filed and published methods and discloses no new algorithmic subject matter. The nine capabilities (soft-vote readout, token-ladder / product-quantizer, drift/noise co-channels, inference-time fusion, guarded recalibration, label-free self-twin monitor, selective-fidelity and channel-attribution) are the program's filed and deposited methods, and the separability screen is previously published; k-means and vector quantization, Fisher discriminant analysis, nearest-centroid and k-nearest-neighbor classification, random forests, product quantization, a distance-based out-of-distribution detector, a two-sample shift detector, pretrained time-series foundation models used only as comparators, and principal-component analysis are established prior art or standard tools, used only as tools or comparators. The method characterized — the class-discriminant single-token codebook encoder, its composable capabilities, and its label-free nearest-centroid novelty score (the filed self-twin / Tier-0 gate), together with the previously published separability screen — is the subject of filed and pending U.S. patent applications held by the authors, or is already published. The results reported here — that the composition is non-destructive and deterministic, that one separability screen predicts the composed safety gate while being blind to corruption-robustness, that label-free recentering recovers a minority of cross-domain loss, and that the stack runs as a bounded, deterministic, sub-milliwatt computation on commodity microcontrollers — are properties of those filed and published methods composed and stressed, measured on public data and on named development hardware, not new subject matter. The per-deployment configuration-selection and input-conditioning procedure, and the silicon productization and commissioning toolchain, are retained as trade secrets and deliberately not disclosed; the platform is characterized here with the plain labeled capabilities, the published screen, and the measured envelope. No third-party partner, customer, or commercial collaborator is named; the development kit and microcontroller named for the on-silicon measurements are a public hardware platform cited for reproducibility, not a partner. The capabilities are characterized for machine-health and sensing tasks. © 2026 Fieldstone Analytics, LLC and the authors. Inquiries: randolphf@fieldstoneanalyticsllc.com. Companion deposits (spiral-domain-encoder-campaign) · "The Frozen Token" capstone — the four-laws / competence-zone theory this study belongs to (doi:10.5281/zenodo.23002239). Is part of. · When Does a Frozen One-Byte Token Suffice? — the published separability screen this study extends to the composed stack (doi:10.5281/zenodo.23288053). · The Competence-Boundary Interlock — the single-interlock axis-specific reach this study replicates under composition (doi:10.5281/zenodo.23347775). · Watch the Input, Not the Representation — the self-twin monitor as an AI-runtime gate (doi:10.5281/zenodo.23351878). · The channel-partition token ladder — the accuracy capability (doi:10.5281/zenodo.22003179). · Soft-vote readout of a frozen single-token codebook (doi:10.5281/zenodo.22866039). · The Cheapest Digital Twin — label-free self-twin (doi:10.5281/zenodo.22922678). · The Service-Life Trajectory — drift / re-commission (doi:10.5281/zenodo.23321165). · The Deployment-Sizing Law (doi:10.5281/zenodo.23323323). · Class-discriminant single-token codebook — base encoder (doi:10.5281/zenodo.20788187). Keywords composable platform; multi-capability composition; non-interference; cross-domain stress; distribution shift; corruption robustness; out-of-distribution detection; safety gate; self-twin; Tier-0 gate; separability law; separability screen; kNN purity; competence zone; commissioning; pre-deployment screen; deployment de-risking; axis-specific reach; honest negatives; pre-registration; permutation test; rep