OBAN k-flow: code and results for exact minimum sound and complete evidence certificates in retrieval-augmented generation

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Code and complete experimental outputs for the manuscript "OBAN k-flow: Exact minimum sound and complete evidence certificates for retrieval-augmented generation" (V. Oban, submitted to Information Processing & Management). OBAN k-flow selects the smallest set of retrieved passages that is both sound (every atomic claim of a fixed answer is supported through an admissible witness hypergraph) and complete (every relevant omitted passage is dominated by the selected ones in a capacitated maximum-flow network), and returns a dual-bound proof of minimality from a single mixed-integer program. The package contains: code/: certification engine (witness hypergraph, two-sided MILP with SciPy/HiGHS, independent max-flow recheck), the SciFact and HotpotQA pipeline (BM25, dense, hybrid RRF and cross-encoder retrieval, extractive reader, NLI verification), statistical analysis, report generation, the Kaggle notebooks used for the experiments, and oracle tests against exhaustive enumeration; results/: recorded verifier scores and all certificates for three NLI verifiers (DeBERTa-v3 base and large cross-encoders), with and without the relevance gate, plus per-query analysis outputs; report/: all result tables (Excel), a Word report and the attribution and verifier-comparison summaries. Licence: code under MIT; results under CC BY 4.0. Claims, questions and answers taken from SciFact and HotpotQA remain under the licences of their original releases.