The most central scenario from my perspective is that there is massive competitive pressure and some amount of motivated denial.
It also might be relatively easily to paper over scheming which makes the motivated denial easier. Minimally, just ongoingly training against the examples of misbehaviour you’ve found might remove obvious misalignment.
(Obvious to whom might be an important question here.)
The most central scenario from my perspective is that there is massive competitive pressure and some amount of motivated denial.
It also might be relatively easily to paper over scheming which makes the motivated denial easier. Minimally, just ongoingly training against the examples of misbehaviour you’ve found might remove obvious misalignment.
(Obvious to whom might be an important question here.)