An Audited Open Record of Three Deep-Blue TADF Host Materials with Benchmark Tasks for Machine Learning

We present an audited measurement record of three phosphine-oxide hosts for deep-blue thermally activated delayed fluorescence (TADF), with provenance metadata and benchmark tasks for machine learning. A scripted audit converted a 5276-file archive into 1662 files (1639 instrument files and 23 instrument-side fit tables) in ten record families, identifying 44 files stored under incorrect material or medium labels and retaining parsed settings, provenance tiers and quality flags. Colour recomputed from the spectra agrees with instrument reports, while gradient-boosted trees detect Raman-scattering artefacts in single emission spectra with a grouped ROC AUC of 0.92. Within-host interpolation is feasible, but three materials do not support broad property prediction for unseen hosts; the released scripts, manifest and task tables provide a reproducible baseline and a schema for adding materials.