Calibrated Residual Tube Projection During Conditional Diffusion Sampling for Thermal Time Series — Code and Experimental Evidence

Abstract

Code and experimental evidence supporting Calibrated Residual Tube Projection During Conditional Diffusion Sampling for Thermal Time Series, by Hyojin Park and Jinhong Yang (corresponding author). The release contains the selected B6 checkpoint, original method implementations, exact simulator and score settings, frozen selections, per-context evaluation records, diagnostic analyses, frozen detector coefficients, author-created figures, and a simulator reconstruction fixture. CPU audits verify saved aggregates, paired intervals, event metrics, decision criteria, cohort separation, journal chains, file hashes, and eight reconstructed fit trajectories. The study uses one fixed noiseless two-node thermal simulator and one selected diffusion generator. Repeated correction reduced true-equation residual by 11.07% and 10.53% on two three-stream cohorts, while both failed the original event-preservation and rollback criteria. A separate controlled detector showed lower T2 timing loss with corrected-only than raw-only synthetic training; real-only training retained lower loss. Deterministic ridge regression nearly recovered the simulator response. Original and later evaluation criteria remain distinguished. This is an evidence and code companion, not a self-contained one-command full retraining release. Full generated-trajectory arrays, other adapters' weights, the manuscript, internal authoring files, and AI-Hub source data are excluded. Historical runners retain original paths and execution gates. The measured-battery branch is represented by derived applicability counts only; no real-battery warning efficacy or broad physical generalization is claimed. License scope: original Python source in dt_diffpinn/ and tools/ is MIT; original synthetic data, results, model weights, figures, documentation, and metadata are CC BY 4.0. These licenses apply to their respective file categories. Third-party dependencies and AI-Hub data are not redistributed or relicensed. No journal acceptance or article publication is asserted. Funding: This work was supported by a grant from Research year of Inje University in 2026(20260050).