Using Opus 4.6 for research code assistance in physics/chemistry, I've also found that, in situations where I know I'm right, and I know it has gone down a line of incorrect reasoning and assumptions, it will respond to my corrections by pointing out that I'm obviously right, but if enough of the mistakes are in the context, it will then flip back to working based on them: the exclamations of my being right are just superficial. This is not enormously surprising, based on how LLMs work, but is frustrating.

Short of clearing context, it is difficult to escape from this situation, and worse, the tendency for the model to put explanatory comments in code and writing means that it often writes code, or presents data, that is correct, but then attaches completely bogus scientific babbling to it, which, if not removed, can infect cleared contexts.