tcmay's approach is elegant, and it's refreshing to find a practical use for AI approaches. I think it needs one more step, though. If we learn to quantify our trust in a key, we still need to know what the threshold should be for a given application. Maybe I can get by with a 0.05 Bel for posting to cypherpunks, but maybe I want a 0.95 for a monetary transaction. Seems to need a comprehensive risk management approach,