We study the extension of applicability of system-level testing techniques to the construction of a consistent model of (legacy) systems under test, which are seen as black boxes. We gather observations via an automated test environment and systematically extend available test suites according to learning procedures. Testing plays two roles here: (i) as an application domain and (ii) as the enabling technology for the adopted learning technique. The benefits include enhanced error detection and diagnosis, both during the testing phase and the online test of deployed systems at customer sites.