
A technician tells you your washing machine needs a €275 drain-pump replacement. But as a homeowner, how do you know whether that recommendation came from a real diagnostic trail or a guess? RepairProof audits the evidence behind the repair you are being asked to approve. For our validated Bosch E18 case, RepairProof retrieves official manufacturer guidance through Bright Data, turns the documented fault causes into explicit evidence requirements and separates every piece of information by provenance: manufacturer facts, technician claims, user observations and RepairProof inferences. The result is not a generic AI answer. RepairProof maintains an explicit evidence state and produces one of three outcomes: NOT YET SUPPORTED, SUPPORTED BASED ON SUPPLIED EVIDENCE, or INSUFFICIENT EVIDENCE. The key behavior is that the verdict can change in either direction. Ruling out only one possible cause does not produce approval. When sufficient technician evidence is supplied, the repair becomes supported. If later evidence contradicts that case—for example, a correction that the impeller was actually jammed—the verdict retracts back to NOT YET SUPPORTED. RepairProof does not claim to verify that a technician actually performed a reported test, and it does not coach consumers through hazardous diagnostics. It audits whether the supplied evidence, if accurate, supports the repair. The current hackathon build deliberately implements one proof-grade washing-machine case rather than pretending to support every appliance with weak reasoning. The same evidence-contract architecture can expand to additional fault codes and models, and has potential applications for consumers, warranty providers, insurers and repair networks. Built primarily with native.builder, with live official-source retrieval through Bright Data, secure server-side Edge Functions, a deterministic verdict engine, and 47 passing Vitest and Playwright contract tests. AI finds the evidence. Proof rules decide.
10 Aug 2026