Warrant: Reproducible Evaluation and Evidence Management for AI Detectors

A detector score is not itself evidence that an operating decision satisfies a declared requirement. Warrant is a local Python toolkit connecting scored, labelled inputs to explicit decisions, retained records and arithmetic replay. This report describes its architecture and investigates the scope of its empirical and statistical claims. Retained benchmark results show within-envelope error-label discrimination alongside low recall at a restrictive threshold, and a constructed context-transfer experiment in which a proposed pooling strategy did not remain robust. Other artifacts cover configuration, declared-workload sampling, monitoring simulations and bounded audit studies. A source-connected investigation finds that all 92 quantitative README rows match retained fields across 21 artifacts, while 15 rows point directly to dirty-source records. A separately implemented rank/count oracle agrees on 25 frozen-score targets, directly covering 19 README point claims. A preserved tied score counterexample also limits the general interpretation of the native adaptive returned false-positive bound. These observations support an inspectable evaluation and evidence management artifact, not universal assurance, independently adjudicated labels or production validation.