Version-lag evidence for trust-to-act decisions over global state in decentralized microservices — code and results

Abstract

Code and results for the article Version-lag evidence for trust-to-act decisions over global state in decentralized microservices (H. A. El-Ghareeb). The record holds two archives. nsdsl-software-v3.zip (MIT) — the implementation: the decentralized act/abstain decision layer in which each peer encodes an item's version lag and storage status as a graded (T, I, F) triple, the weighted aggregation operators and score functions, the consistency baselines (strong, eventual, quorum, partial quorum, single peer) and the scalar controls (calibrated probability gate, logit gate, isotonic gate) that see the same observables, the ANTLR-based rules DSL, the discrete-event simulator and workload generator with its attack menu, the trace-driven workload loaders (UCI Online Retail II; Twitter cache trace), the FastAPI microservice testbed with its Docker Compose deployments (localhost, plain Docker, emulated WAN with Toxiproxy and per-packet loss), the latency and throughput benches, the metaheuristic weight tuning, the single canonical aggregator, the table and figure generators, the test suite (187 tests), REPRODUCE.md, CITATION.cff and LICENSE. nsdsl-data-v3.zip (CC BY 4.0) — the two content-hashed parameter files (config/calibration.json, config/mej.json), the per-trial rows of the headline grid (10 seeds, 100 repetitions per cell, 400 decisions per trial) and of the two trace-driven workloads, every aggregate the article reports (per-configuration summary with Holm-corrected pairwise tests and Cohen's d, multi-seed validation, unconditional stale rate, risk–coverage curves and the matched-availability comparison, scalar-control calibration, thresholds, unseen-workload holdout, encoding sensitivity, the representation × operator × score factorial with its ANOVA, budget-normalised variants, adversarial attack menu, transfer, effect sizes in engineering units, the lag-isolation ablation, Byzantine robustness, scalability, sensitivity, convergence, throughput, weight tuning), the measured latency on the localhost, Docker and emulated-WAN testbeds including the three back-to-back repeats with their host-load samples, the generated LaTeX tables and every figure, the environment lock and a DATA_README.md. Every reported number regenerates from the software, the parameter files and the recorded seeds; the SHA-256 of each parameter file is stamped into every result CSV header and the aggregator refuses to mix files that disagree. The trace-driven workloads use two public traces (CC BY 4.0) fetched by the shipped scripts; no other third-party data is involved. The concept DOI 10.5281/zenodo.20932988 always resolves to the latest version.