Beyond Node Counts: Correlated-Failure Audits for Multi-Role Permissioned Ledgers

Reproducibility artifact for “Beyond Node Counts: Correlated-Failure Audits for Multi-Role Permissioned Ledgers” Version 1.0.0 accompanies the ETECOM 2026 paper by Son Huu Trung Dang, An Van Nguyen, and Tam Huu Tran. Beyond Node Counts examines how shared dependencies affect threshold availability in permissioned-ledger deployments. Given a fixed component placement, an inventory of disruptive actions, positive integer action weights, and role-specific thresholds, the implementation computes minimum-weight disruption witnesses for validator outage, committee outage, pipeline OR, and simultaneous AND. The model describes each action through the components it disables and combines selected actions by set union. The artifact implements the mixed-integer linear formulation, exhaustive integer enumeration, a quota-cover greedy baseline, and independent bounded checks. Numerical acceptance checks validate returned witnesses and objective values; a bounded exact fallback handles specified numerical discrepancies. This record contains the executable implementation and the experimental material supporting the paper: constructed configurations, benchmark inputs, reference results, sampled trace mappings, behavioral tests, and recorded ledger and dependency experiments. FILES - etecom-2026-artifact-1.0.0.zip: the execution archive, including Docker setup, Python source, experiment runners, input data, reference outputs, and supplementary evidence.- SHA256SUMS: SHA-256 checksums for the prepared release files. ARTIFACT_MANIFEST.json records the delivered inventory and individual file hashes. EVIDENCE_PROVENANCE.json identifies the selected historical evidence, its original hashes, and the entries omitted from this execution package. HOW TO RUN Extract the archive and run these commands from its root directory: docker build -t etecom-artifact:1.0.0 . docker run --rm --network=none etecom-artifact:1.0.0 The build requires Internet access to obtain the Python image and dependencies. Experiment execution then runs offline with networking disabled. Docker provides the Python environment; no local scientific packages or LaTeX installation are required. The container checks the delivery manifest, executes the test suite, recomputes the primary evaluation, and verifies the saved supplementary evidence. Successful execution ends with: experiment reproduction passed Results appear in /artifact/evaluation/generated/ inside the container, with execution logs in /artifact/build_logs/. The README provides commands for retaining a container and copying these outputs to the host. The archive also includes reference outputs for inspection. For native execution, use Python 3.11, 3.12, or 3.13. On Windows, run setup.bat followed by run.bat; on Linux, run sh setup.sh followed by sh run.sh. Setup installs the dependencies specified in requirements.lock. The Docker workflow has been tested from the extracted release archive on Linux containers with networking disabled. DATASETS The artifact uses both Google ClusterData 2011 and Microsoft 1999 desktop reachability data. The Google machine-event shard resides under data/external/google_clusterdata_2011/. The default workflow recomputes 100,000 paired host mappings and compares 800,000 outage durations against an independent interval oracle. The Microsoft material resides under supplementary/microsoft1999/. It includes 100,000 saved mappings, replay summaries, statistical checks, source provenance, and two independent duration-oracle implementations. The default workflow verifies the saved records. Full replay from source requires the supplied download script, which checks pinned hashes before extracting the input. The archive omits the raw Microsoft trace because the recorded provenance review did not establish permission for downstream redistribution. WHAT THE

Results

COVER The default workflow passes 126 tests. Computational checks include 720 MILP-to-enumeration objective comparisons, 599 bounded Set Cover reduction cases, and 179,520 bounded quota instances. Constructed experiments examine inventory omissions, incidence uncertainty, action-weight sensitivity, approximation behavior, and runtime scaling. Supplementary numerical stress records cover 2,000 small instances, with zero accepted mismatches and 100 exact fallbacks. A separate script supports rerunning this stress experiment. The recorded CometBFT experiment tests quorum conformance under controlled validator-process pauses on one host. The prospective dependency experiment examines a synthetic authorization path through 56 primary relay windows, with original controls and separately recorded corrective controls. Offline verification checks the underlying observations; the default workflow does not launch new live ledger experiments.

Interpretation

The historical traces support counterfactual ledger-role assignments and conditional availability comparisons. They do not supply observed ledger-role inventories, causal failure domains, or calibrated disruption costs. The dependency experiment identifies configurations that violate the union-incidence assumption. Its targeted route change matches a simple placement heuristic and does not improve the mean outage count of the tested matrix. The reported evidence supports the stated model checks and experimental comparisons, without establishing production effectiveness or optimizer superiority.

Funding

AND ATTRIBUTION Vietnamese-German University (VGU) funded this research. The authors acknowledge ChatGPT assistance with evaluation and plotting code and take responsibility for the final implementation. Licence: MIT for authored software. Google inputs retain CC BY 4.0 attribution, and supplementary materials retain their applicable source notices. See LICENSE, THIRD_PARTY_NOTICES.md, and the dataset-specific notices inside the archive.