LLM-Explainable Consumer SSD Health Scoring with SMART Attributes and Reliability Statistics

Abstract

Reliable consumer SSD and NVMe assessment requires a model that distinguishes observed SMART evidence from population context and communicates risk without presenting a diagnostic certainty that the data cannot support. This paper presents LESHS, an explainable health-scoring framework built on the LinuxHW SMART Repository. The analysis processed 162,383 rows from the All_SSD.md and All_NVMe.md appendices, removed one exact duplicate, and retained 162,382 records: 81,660 SATA SSD samples and 80,722 NVMe samples. The repository-defined important-error flag was positive for 5.09% of retained records, with rates of 8.64% for SSD and 1.49% for NVMe. A strict non-leakage experiment withheld each record's Err and MTBF fields and predicted the important-error flag from drive type, power-on exposure, capacity, and training-set manufacturer, model, and capacity-bucket priors. On a stratified test set of 24,358 records, the calibrated LESHS model achieved ROC-AUC 0.8873, PR-AUC 0.3892, F1 0.4066, and Brier score 0.0390. A separate consistency analysis reconstructed the LinuxHW MTBF statistic with MAE 0.00255 years and R2 0.99999. Operational scoring then combined the calibrated prior with the observed important-error count and an explicit penalty for short observation windows. A constrained explanation schema translated the resulting evidence into concise statements about error presence, power-on duration, MTBF, and score band. The findings show that calibrated population priors and direct SMART evidence play complementary roles in consumer-drive health assessment.