Reproducibility package for "Auditing autonomous water-allocation decisions under scarcity: a reproducible challenge benchmark for value-sensitive governance"
Abstract
This repository contains the complete reproducibility package accompanying the manuscript “Auditing autonomous water-allocation decisions under scarcity: a reproducible challenge benchmark for value-sensitive governance”. The release implements a synthetic single-reservoir benchmark for testing technical auditability, reviewability, and normative sensitivity in autonomous water-allocation decisions under scarcity. The package includes the complete Python source code, pinned software dependencies, configuration files, an end-to-end reproduction script, the complete decision log, per-seed operational metrics, statistical outputs, audit-task validation, controlled fault-injection experiments, sensitivity analyses, figure source files, and a frozen SHA-256 reference manifest. The benchmark uses 1,000 matched simulation seeds and compares a fixed reference policy with an adaptive declared-priority policy under a drought shock. Three nested evidence levels are evaluated: outcome-only records, operational traces, and full reviewable traces. The audit challenge includes 2,100 controlled corruptions across seven fault classes together with 300 clean controls. Sensitivity analyses examine stakeholder-order conventions, drought severity, initial governance-priority states, and alternative reporting-weight regimes. All data in the archive are synthetic. The benchmark is intended to test whether retained decision evidence supports independent reconstruction, procedural verification, outcome recomputation, provenance tracing, and fault detection. It is not a calibrated water-management model, a deployable allocation policy, or a claim of universal ethical correctness or algorithmic superiority. The full analytical workflow can be regenerated using the supplied run_all.sh script. The reproduction workflow does not rewrite the frozen reference checksum manifest, allowing regenerated deterministic outputs and tracked source files to be independently checked against the archived reference.
