I would never call that extra understanding just “slightly” less random. For example going from “random” to “understands what a pass alive group is” is HUGE to me. Understanding the very basic capture rule is already a huge step from random. And that does not even require any notion of eyes at all, just “count liberties, play adjacent to enemy group having less liberties in total” for example.
It is relatively simple to code to forbid suiciding a pass alive group (if you do allow it to suicide when in seki, at least, as detecting seki is harder). Just add the rule that if you have any pass-alive group (which can be checked mechanically with Benson’s algorithm) , then a move that turns a pass-alive stone into a no longer pass alive stone is forbidden. Then just play any (other) random move, until only forbidden moves remain, then pass. This bot would fill it’s own territory, but if it has a group with two very well defined eyes, it will not fill them. It will however fill it’s own eyes and kill itself many times, when the group was alive but not yet pass-alive. That is yet another much more powerful level of understanding, as it is not just a “simple fast check” as benson algorithm, but rather requires tsumego-solving (even if only very simple tsumego like the very basic elementary shapes a beginner learns on day 1 and 2 of experimenting).
It would be an interesting experiment to see what the win% between random and “random but does not suicide pass-alive groups” is (is it 90%? 99%? 99,9999%?). It will completely crush a random bot, just look at the previous game where both players have pass-alive groups near the end, but white for example suicides a large group by self-filling.