Paper — Supplementary materials: 480 latency measurements + analysis pipeline + 60 anonymised LFW embedding hashes

This archive contains the full reproducibility package for the paper "Infrastructure Sizing and Scalability Analysis of a Web-Based Face Recognition Attendance System: An Empirical Study" A 4×4 between-subjects factorial study on a web-based face-recognition attendance system, varying VM tier (4 levels from 2 to 8 vCPU on commodity Ubuntu 22.04 + InsightFace buffalo_sc) and enrolled-class size (10, 20, 30, 50 students), with 30 trials per cell (480 total measurements). All 480 trials matched the expected identity (TPR = 100% by construction). All 16 cells passed the pre-specified P95 ≤ 3000 ms acceptance threshold by an order of magnitude (max observed P95 = 202 ms). Contents: - 480-row latency dataset (`data/raw/all.csv`, SHA-256 41df7f5c8ab97e9ffd64a688b1111dc716db4006f26c55afae7d3de3267c0fce) plus 16 per-cell views. - 8 processed-analysis CSVs (descriptive statistics, Shapiro-Wilk normality, Kruskal-Wallis, log-transform two-way ANOVA, Tukey HSD via match regression, H4 OLS regression per tier, H5 Pearson correlation, sensitivity analyses including outlier-exclusion and Alexander-Govern robust ANOVA). - 6 deterministic Python scripts (analysis pipeline, figure generator, benchmark client, LFW-identity sampler, probe preprocessor, enrollment script). - 8 publication-quality figures (PNG + TIFF, 300 DPI, MDPI-sized). - SHA-256 manifest of 60 anonymised LFW embeddings (`identity_001…060`) for verification. - 4 supporting documents (reproducibility manifest, locked experimental protocol, dataset provenance, analysis report). Three reproduction levels are documented in `README-supplementary.md`: - L1 (analysis): re-run the statistical pipeline against the published dataset (~2 minutes, Python only). - L2 (embedding regeneration): re-derive the 60 anonymised embeddings from LFW (download separately from UMass). - L3 (full benchmark replay): re-collect the 480 measurements via the four cams-A{1..4} VirtualBox VMs. The hypotheses, sample size, stopping rule, and analysis pipeline were pre-registered on the Open Science Framework before any analysis was performed; OSF project page https://osf.io/9kb6a/ (registration DOI in the related-identifiers section once locked). This deposit does NOT redistribute LFW images (download separately from http://vis-www.cs.umass.edu/lfw/) and does NOT publish the LFW name → anonymised-index mapping (preserves LFW identity privacy). The anonymised embedding hashes allow third-party verification of regenerated embeddings without exposing identity information. Code is licensed MIT (`LICENSE-code`); data + figures + manifests are licensed CC-BY-4.0 (`LICENSE-data`).