In rule-based reasoning, a conclusion often rests on a chain of evidence drawn from external sources. But that evidence is not permanent: a source may be withdrawn, and when it is, the entire justification can disappear. A new paper on arXiv examines this fragility for Horn reasoning, where conclusions are derived from a set of rules and facts.
The authors point out a subtle complication: having multiple derivations of the same conclusion does not guarantee robustness. Different derivations may rely on the same source, so a single failure can knock out several lines of support at once. This correlation between failures makes the problem of choosing reliable evidence more difficult than simply collecting redundant proofs.
The paper introduces the notion of minimum-cost robust evidence certificates—sets of sources that keep the conclusion valid even when some sources fail. By modeling correlated failures explicitly, the work aims to provide a principled way to select evidence that is both cheap and resilient. The full details of the proposed approach are not yet available in the abstract, but the problem it addresses is central to any system that depends on external, potentially unreliable information.