Federated Multi-Institutional H-Pipeline for Privacy-Preserving Cloud-LLM Clinical Reasoning Across Healthcare Networks: A 255-Study Validation Campaign of Stages F- 1 through F-5 with Executed Adversarial Audit and Composed-Mitigation Specification

v2 (2026-06): Corrected attribution-- compressibility and privacy are now attributed to the codebook-quantization layer, lossless invertibility to the spiral front-end. All numerical results unchanged. This deposit accompanies the v1.0 manuscript for Paper 13 of the spiral-domain encoder validation campaign series. It contains the complete pre-registration corpus, runner code, summary verdicts, per-study reports, figures, POC addenda, and arXiv-format manuscript for a 255-study empirical validation campaign of the federated multi-institutional extension of the H-pipeline architecture introduced in Paper 12. The campaign comprises 215 pre-registered hypothesis-testing decision rules across eleven thematic sections (§1–§11) plus 40 post-hoc gap-closure / closeout / mitigation-validation addenda across three additional sections (§12, §13, §14), executed in a continuous deterministic protocol (PYTHONHASHSEED=0) with zero post-hoc threshold adjustments to the pre-registered §1–§11 verdicts. The campaign achieves 84.3% SUPPORTED (215 of 255 hypotheses) across the full §1–§14 envelope. Pre-registered verdicts are honestly reported in SUPPORTED / NOT_SUPPORTED classes; the §12 executed-attack findings substantively revise §4 placeholder adversarial values; bounded-negative findings are preserved as deployment-boundary characterizations rather than suppressed. Headline findings · Token-reduction utility preserved under federation: 17,999× on real PhysioNet MIT-BIH at N=5 institutions (§12-B §949). · All four Slide-8 mathematical properties of the spiral-domain encoder preserved under N=2 federation on real data (§12-B §952). · Five byzantine-robust aggregation rules characterized (geometric median, Krum, trimmed mean confirmed; concatenated, k-means-of-centroids, quorum-weighted compared). · Critical empirical-vs-theoretical privacy gap surfaced: executed capable-adversary MIA AUC = 0.97 on real PhysioNet MIT-BIH (§12-B §950) vs §4 §796 placeholder 0.55. The (ε_fed=5.0)-DP theoretical bound is probabilistic, not empirical, against identity-bearing morphological features. · Centroid-reconstruction DLG attack: MSE = 0.0006 (§12-C §957) — federated codebook centroids are reconstructive of training samples. · §14 mitigation validation: σ tuning alone is insufficient at quorum k≥2 — k=1 + σ=2.5 is the §979 verified-working configuration (TPR=0.993, FPR=0.000). Single-direction mitigations (identity-axis projection, codebook centroid noise, tighter ε) show large directional effects (|Cohen's d| > 6) but none individually closes the empirical privacy gap. Related provisional patent applications The Phase XXVI campaign establishes the empirical foundation for three new standalone U.S. provisional patent applications by the inventors, all filed at USPTO on 2026-06-11: Parent I (64/084,807 — federated multi-institutional H-pipeline), Parent J (64/084,817 — composed codebook-centroid privacy hardening), and Parent K (64/084,821 — pre-encoder identity-axis projection). Each is a standalone provisional with an independent priority date; 12-month non-provisional conversion deadline is 2027-06-11. Companion deposits in the campaign series · Paper 8 — Phase XXI musculoskeletal-DT validation: 10.5281/zenodo.20466035 · Paper 9 — Phase XXII surgical-robot patient-pair DT: 10.5281/zenodo.20470763 · Paper 10 — Phase XXIII unidirectional release-only bioelectric DT: 10.5281/zenodo.20480183 · Paper 11 — Phase XXIV bidirectional closed-loop bioelectric DT: 10.5281/zenodo.20495976 · Paper 12 — Phase XXV IoT-LLM pipeline: 10.5281/zenodo.20499453 Reproducibility All 255 studies execute deterministically under PYTHONHASHSEED=0. The reproducibility corpus includes the federated H-pipeline harness library, 14 section runner scripts, 255 per-study summary JSONs, and 255 per-study REPORT_NNN.md files. End-to-end campaign reproduction wall-clock: ~10 minutes on Apple M2 / 16 GB. License Creative Commons Attribution 4.0 International (CC-BY 4.0). Commercial deployment of the federated H-pipeline architecture requires a license to Fieldstone Analytics, LLC. Contact: [email protected]. Citation Ferlic, R.J. and Ferlic, K.K. (2026). Federated Multi-Institutional H-Pipeline for Privacy-Preserving Cloud-LLM Clinical Reasoning Across Healthcare Networks. Zenodo. https://doi.org/10.5281/zenodo.20595834