AI can be right and yet fragile, if it doesn’t know when its confidence was misplaced.

I propose the Tension Principle: measure the gap between a model’s predicted prediction accuracy (PPA) and its actual prediction accuracy (APA), defined as: 

T = |PPA − APA| (with extensions to prevent gaming)

This gives the model a signal about its own epistemic reliability. Example: a chatbot expects 95% accuracy on a hard question but scores 50%. Tension flags that mismatch — not for being wrong, but for being wrongly sure.

This second-order signal could:
• Detect overconfidence or hesitation even in correct answers
• Catch slow calibration drift before behavioral issues emerge
• Add an internal self-correction layer to complement RLHF

APA isn’t always directly observable — proxies or approximations may be needed. But the principle is simple: self-monitor your own trust in yourself.

Does this hold water technically? Has anyone explored similar second-order miscalibration signals? 

Details here: On the Principle of Tension in Self-Regulating Systems