Stochastic Workload Synthesis via Constrained Generative AI: > Enhancing Resilience Diagnostics in Containerized Microservices

This deposit contains all benchmark data, experimental artefacts, and software code associated with the manuscript: "Stochastic Workload Synthesis via Constrained Generative AI: Enhancing Resilience Diagnostics in Containerized Microservices" Submitted to Applied Sciences (MDPI), Special Issue: Intelligent Systems for Cloud-Native Reliability, 2026. Data files (CSV): all_runs.csv — Complete per-run raw data for all 960 experiment runs (4 conditions × 4 faults × 3 services × 4 profiles × 5 replications): MTTD, precision, recall, F1, RCA result, coverage bucket per run table_entropy_per_run.csv — Coverage entropy (bits) per replication for all 4 conditions table_mttd_summary.csv — MTTD mean, SD, 95% CI, Cliff's delta, Bonferroni-corrected p-values table_precision_recall_f1.csv — Precision, recall, F1 by condition and fault category with 95% CIs table_modality_ablation.csv — Anomaly detection macro F1 for all 7 observability modality combinations table_rca_accuracy.csv — Top-1 RCA accuracy by condition and fault category table_scenario_validity.csv — Scenario generation validity rates at each pipeline stage table_timevae_training_curves.csv — TimeVAE ELBO, reconstruction MSE, KL divergence per epoch for all 5 random seeds table_llm_generation_metadata.csv — Per-run token counts, cost, generation attempts, constraint pass rates Software artefacts: synthchaos_constraint_schema.json — Example constraint specification document used in primary evaluation litmus_cpu_hog_example.yaml — LitmusChaos ChaosExperiment manifest for CPU throttle fault litmus_network_loss_example.yaml — LitmusChaos manifest for network partition fault Testbed: 6-node Kubernetes v1.30.2, Istio 1.21.3, LitmusChaos 3.10.0, Online Boutique v0.9.1, Prometheus 2.52.0, Loki 3.0.0, Jaeger 1.57.0. LLM: Claude Sonnet 4.6. TimeVAE trained on NVIDIA A100 40 GB. License: CC BY 4.0