The journal version of the NASA Ames work. A 503-node model of the ADAPT testbed, compiled into an arithmetic circuit, returned its most probable explanation in about a quarter of a millisecond on average, an order of magnitude faster than the alternatives tested.
The full journal treatment of the ADAPT work. The paper sets out a systematic way to represent an electrical power system as a Bayesian network, supported by an easy-to-use specification language, and compiles the network into an arithmetic circuit so that diagnosis is predictable, fast and exact.
The ADAPT model had grown to 503 nodes and 579 edges. Computing the most probable explanation with the compiled circuit took a mean of 0.2625 milliseconds, with a standard deviation of 0.2028 milliseconds. Variable elimination and join tree propagation also did well on ADAPT, but circuit evaluation was an order of magnitude or more faster.
The two challenges the paper is organised around, building the model and reasoning in real time, are the two challenges Manifest is built around. The numbers here are the peer-reviewed record of what compile-once diagnosis looked like on a real power system in 2010.
ADAPT is a ground testbed at NASA Ames, not a flight system.