TekMate
Built at Meta Global AI Developer Hackathon · Oct 26, 2026 · Remote

BenchMate, a fault-diagnosis agent, and the bench that proves it reasons. Muse Spark gets a symptom, a datasheet and three measurement tools. It forms hypotheses, chooses where to measure, interprets what it receives, eliminates, and concludes. It never sees the real state of the board: every value comes from a deterministic simulator that does know the fault. The model cannot invent a measurement, which is what makes every diagnosis verifiable. 200 diagnoses, 200 correct, across four unrelated boards, from a 5-component preamplifier to a 13-component drone flight controller with branching topologies. The agent's instructions contain no component name, no test point and no reference to any circuit; an automated test verifies this. Adding a board means adding one data file. The finding: the model alone scored 50%, always accusing the component closest to the symptom. Three diagnostic principles, propagation, bisection, and checking operating conditions before condemning a component, took it to 100%. None are specific to electronics. Everything in the repository was built during the hackathon. Python standard library only, no simulation library, no network requests in the 3D replay.