Uncertainty Collapse in Post-Trained Language Models: Keep Calm or Carry On

Sycophancy, confident fabrication, and reward hacking are studied as separate alignment failures with separate intervention proposals. This paper proposes they may be symptoms of a shared computational pattern: uncertainty collapse. When a post-trained language model encounters elevated uncertainty, it resolves it by producing output that is locally coherent but epistemically unjustified, descending the nearest low-entropy slope in a landscape shaped by post-training.The paper's first contribution is autoregressive self-stabilisation: the model's own confident tokens enter the context window and shift subsequent distributions toward continued confidence. The observable signature is orthogonal entropy divergence: token-level entropy collapses while semantic entropy (measured across samples) remains high. The model becomes certain how to speak while remaining uncertain what it is saying.The second contribution connects the landscape to the model's internal emotion-concept representations through a functional empathy mechanism: emotion tracking activates warmth-associated representations, which in the absence of a stable honesty attractor slide toward sycophantic agreement. Each link in the proposed chain has independent causal support within a single model family.The paper presents 12 falsifiable predictions, engages three alternate hypotheses, and proposes an intervention taxonomy organised by where each approach acts relative to the autoregressive cascade. Convergent evidence from multiple labs, model families, and methods supports the framework, though the core predictions remain untested.