Definition
Plain language
When an AI's certainty actually builds up gradually as it works through the problem.
As stated in the literature
The healthy counterpart to premature confidence: mid-chain probes show monotonically rising confidence in the final answer across the trajectory, indicating reasoning that genuinely narrows the model's hypothesis space.
Why it matters: Gradually-built confidence is a signature of reasoning that actually does work, which is what makes chains of thought trustworthy.
For example, the model starts at 30 percent confidence in answer 'B', rises to 60 percent after listing constraints, and reaches 95 percent only after the final check.