VARF: Synthetic Audiogram Generator, Experiment Code and Systematic Review Materials

Supporting materials for the manuscript "VARF: A Five-Layer AI Architecture for Virtual Aural Rehabilitation: Design and Algorithmic Feasibility", submitted to Engineering, Technology & Applied Science Research. Contents. A synthetic audiogram generator (deterministic, seed 20260301) and the complete code for four experiments: a hybrid rule-plus-gradient-boosted triage classifier, adversarial robustness checks, differentially private federated averaging with Renyi differential-privacy accounting, and a LinUCB adaptive difficulty controller benchmarked against a clinical staircase procedure. Also included: the generated dataset of 5,000 cases, a reward-weight sensitivity analysis, a privacy-accountant configuration scan, the systematic review materials (database-specific search strings for PubMed, Scopus, CINAHL and IEEE Xplore with all deviations recorded, and the PRISMA 2020 flow), and the six figures as published. Provenance. code/varf_reproduce.py is the authoritative specification of the method: every numerical value in Section III of the manuscript is its output. PROVENANCE.md records which reported values changed relative to the originally submitted version, and why. Important caveats. All results derive from synthetic data. No record-level patient data and no record-level NHANES data were used at any stage; the generator's parameters were calibrated to published summary statistics only, namely age- and frequency-stratified mean thresholds and standard deviations from the NHANES audiometry examination and the WHO hearing-impairment grade cut-offs. Because the triage labels and the inference-time rule set derive from the same clinical criteria that parameterised the generator, the reported metrics index recoverability of a deliberately encoded structure and are an upper bound on what the pipeline could achieve on real audiograms, not an estimate of it. No clinical validity is claimed. One result is negative. The LinUCB difficulty controller is outperformed by a transformed two-down-one-up staircase, the adaptive procedure clinical speech-in-noise testing already uses (0.767 against 0.703 mean reward; 90.6% against 52.1% of trials within the target zone). The manuscript reports this as a negative result and frames the corresponding architectural layer as an interface specification rather than as evidence for the bandit. Headline results. Triage classifier 0.938 +/- 0.008 accuracy, 0.971 macro-AUC, expected calibration error 0.038 under 5-fold cross-validation. Differentially private federated learning 91.0% accuracy at epsilon approximately 4.23 and 88.1% at epsilon approximately 1.44, against a non-private centralised baseline of 93.5%. Generative AI. The code was written with the assistance of a generative AI tool (Claude, Anthropic). This constitutes use in the research process rather than language editing and is disclosed as such in the manuscript's AI declaration. Reproducing. ./run_all.sh, approximately four minutes on a commodity CPU. Output is deterministic and has been verified identical on two independent machines. Note on completeness. The four per-study review files under review/ (included_studies.csv, data_extraction.csv, quality_assessment.csv, table2_coding_sheet.csv) are correctly structured but not yet populated.