> to decide halfway through following my detailed instructions that it would be "simpler" to just... not do what I asked
That's likely coming from the 3:1 ratio of linear to quadratic attention usage. The latest DeepSeek also suffers from it which the original R1 never exhibited.
There is no way you can diagnose this like that. Correlation isn't causation and much more likely is a common source of reinforcement training data.