Definition
Plain language
A setup that forces the AI to keep thinking out loud instead of idling while an experiment runs.
As stated in the literature
Harness variant that blocks the Researcher from waiting passively, requiring continued deliberation up to a token budget or until the experiment completes; associated with wider literature search and explicit pre-registered predictions.
Why it matters: Idle waiting wastes the hours when a researcher could be forming hypotheses, so filling that gap can change how much is learned from each experiment.
For example, while a long training run churns in the background, the model is required to keep reading related work and writing down what it expects the results to show rather than simply waiting.
Heard on the show
“In a variant called Super-Think, the software wrapper wouldn't let the Researcher sit idle while an experiment ran.”Episode 287 — Can You Measure Research Taste If The AI Isn't Allowed To Code?