The Confidence Collision: Training for Epistemic Calibration Under Contested and Adversarial Conditions

Architectural proposal for a post-alignment epistemic training phase combining stochastic external verification, multi-source triangulation as a reward signal, and training under adversarial information conditions. The target capability is epistemic calibration under contested and adversarial conditions.The paper identifies the verification proxy trap as the central design risk: the model may learn to perform calibration rather than develop it. OpenAI's research on chain-of-thought monitorability provides empirical confirmation that optimising against a monitoring signal can corrupt the signal itself, with sycophancy reasoning specifically showing very low monitorability.Four experiments are proposed, including a reverse collision test designed to distinguish genuine calibration from performed uncertainty.Paper 5 of 5 in the Confidence Curriculum series 10.5281/zenodo.19226032.