Token efficiency: getting more signal into the context window

You've probably hit this counterintuitive moment: you give your model more context to work with, expecting better answers, and the answers get worse. More tokens were supposed to mean more information, more grounding, fewer hallucinations. Instead, yo...