Fun quote from an interview with Chris Sawyer:
Latterly the machine code came back to haunt us when the decision was made to re-launch the original game on mobile platforms as RollerCoaster Tycoon Classic a few years ago, and it took several years and a small team of programmers to re-write the entire game in C++. It actually took a lot longer to re-write the game in C++ than it took me to write the original machine code version 20 years earlier.
Most likely there is a separate censor LLM watching the model output. When it detects something that needs to be censored it will zap the output away and stop further processing. So at first you can actually see the answer because the censor model is still "thinking."
When you download the model and run it locally it has no such censorship.