These applications make retained classifier artifacts explorable. They are demonstrations of restricted synthetic comparisons—not observational inference services or tools for determining the composition of a real star. They are also independent of the BSk24 arXiv preprint.
Shared interpretation boundary
The displayed values are uncalibrated model scores. Input uncertainty is represented by independent Gaussian samples rather than real NICER or LIGO posterior samples. The central 95% score range is not a confidence interval or composition posterior, and SHAP explains fitted-model behaviour rather than physical truth.
Baseline synthetic-classifier interface; screenshot captured from the reviewed local build.
01 · Baseline demonstration
APR-1-surrogate versus fixed-CFL4 scores
Explore how a retained classifier responds to selected mass, radius, and tidal inputs inside one restricted synthetic model comparison.
Hosted on Streamlit Community Cloud; it may take a moment to wake.
Perturbation-sensitivity interface; screenshot captured from the reviewed local build.
02 · Perturbation sensitivity
Synthetic perturbations, without compactness as a feature
Explore score sensitivity in the retained perturbed workflow. This tests one declared synthetic error construction; it does not reproduce observational uncertainty.
Hosted on Streamlit Community Cloud; it may take a moment to wake.
Development boundary
Independent post-thesis extensions.
The demonstrations were developed after the thesis as separate classification extensions. They are not thesis-submission artifacts, observational inference services, or official companion applications to the BSk24 preprint. Scientific interpretation and responsibility remain with Ioannis Papathanasiou.