Finite-Frame Transfer Compatibility: Exact Spectral-Mixing Residuals under Known Circular LTI Transfer Finite spectral estimation and known linear transfer do not generally commute. This work shows that the resulting spectral discrepancy is not merely an uncontrolled finite-window artifact: under a known circular LTI transfer and an exact all-origin finite-frame mixing kernel, the compatibility residual is analytically defined and quantitatively computable without fitted calibration. For an input power spectrum S, a nonnegative finite-frame mixing operator K, and power transfer a=∣H∣2, the fitted log-frequency slope of the compatibility defect is exactly the residual between ideal transfer slope and observed spectral-slope migration on a fixed frequency mask. Its pointwise depth curvature is also determined by a variance of log transfer gain under a depth-tilted spectral measure. The frozen benchmark contains 64 synthetic records and 65 nonoverlapping real-data blocks from electrocardiography, Bitcoin minute returns, and solar-wind magnetic-field data. Across 3483 retained cells, pooled R2 for ΔR4=R4−R3 ranges from 0.9891 to 0.9994 across the nine real dataset–estimator groups. After removing fixed-configuration means, R2 remains 0.6608–0.9943; across 81 fixed real configurations, the median R2 is 0.8782. The record includes the manuscript, frozen data, executed publication run, source code, dependency specification, provenance and licensing documentation, and SHA-256 manifests required to reproduce and audit the reported results. Reproducibility DOI: 10.5281/zenodo.22056027Corresponding author: gmtheory@outlook.fr Licensing is file- and source-specific. See DATA_LICENSES_AND_ATTRIBUTION.md for upstream licenses, attribution requirements, and provenance of the redistributed data.
Compact AI systems make local language-model experimentation increasingly accessible, yet practical evidence for multi-node training on desktop-class accelerators remains limited. This report presents a proof-of-concept deployment of distributed NanoChat pretraining across two NVIDIA DGX Spark systems, each with a GB10 Grace Blackwell system-on-chip and 128 GB of unified memory, administered remotely over a Tailscale mesh VPN and connected for training by a dedicated 200 Gb/s QSFP56 direct fiber link. PyTorch torchrun, DDP, and NCCL were configured with one process per node, a depth-20 NanoChat model, a local batch size of 32 per node, and a 2,048-token context, giving a global batch of 131,072 tokens per step. The run sustained a step time of about 69.4 s (about 1,890 tokens/s), processing about 653 million tokens over four days. We document link configuration, container setup, interface binding, a step-zero evaluation bug that triggered NCCL timeouts, checkpointing, and troubleshooting lessons, as a reproducibility reference for small labs. We also built a cybersecurity fine-tuning dataset from 77 CISA advisories (338 training, 37 validation conversations) and ran a 17-question held-out evaluation comparing a baseline SFT checkpoint against a CTI-augmented checkpoint with an Ollama-hosted LLM judge. CTI-specific categories improved while general-knowledge categories regressed, for a small overall change from 2.06 to 2.29 on a 0-10 scale. The same cluster supports a 400-level AI course (CS 426) and a query engine for CompTIA Security+ POGIL activities in CBS 255, showing modest local infrastructure can serve both research and teaching. The study establishes feasibility rather than a scaling-efficiency claim, since single-node throughput used for comparison was estimated, not measured under matched conditions. Runbook and scripts are available (see Code Availability).