What I tell people about my experiences with Versu:
You want your creation to surprise you; but you don’t want it to surprise you in a surprising way. The second derivative of surprise should be 0.
Here’s what I mean by that. In testing for Versu, we had cool things happen like:
– recombining characters from different settings, we allowed Patrick (a parodic modern-day frat bro character) to chat with Admiral Heron (an early 19th century British admiral). Patrick started talking about his success with the ladies, and Admiral Heron responded by talking about how his sailors caught STDs while visiting the Pacific. In context, it felt as though Heron was perhaps reprimanding Patrick for his behavior or at least drawing a comparison with his sailors. Procedurally, all that was happening was that each of those characters had some smalltalk dialogue associated with the conversation topic “sex”, and this caused an automatic transition here.
– locations and props in locations could provide affordances for expressing particular behavior. In one case, two characters got into an argument while standing in a park, and then one of the characters pushed the other into a lake – because the lake provided the affordance of pushing someone into it as an aggressive action, and the aggressive behavior became available because of the angry mood.
Those were both things that were not specifically intended by the author and both produced output that was funny and plausible. They were also the types of surprise that we expected to be facilitating when we designed the system the way we did.
But we also had loads of things happen that were outcomes of the system that were hard to understand or justify narratively. These were the surprising surprises, and they were undesirable. For instance
– a case where a character was rude and aggressive to the protagonist at the first meeting, even though they’d never interacted before. Tracing back through the simulation history, it turned out that that character had exchanged gossip with another NPC offstage and had heard negative things about the protagonist. Here the simulation was still working (in a sense) as intended, but the problem was that the system wasn’t designed to show the player that that was what had happened, so instead of being an instance of persuasive, interesting NPC autonomy, it was a thing that looked like a bug.
– cases where, for a combination of reasons, characters became avoidant and hid off in the corners of the map; or became aggressive and attacked one another; or hit on their own relatives; or had sex in front of coworkers during an office meeting; or various other things of that nature. These were cases where the simulation was not working as intended, either because the rules we had written had unanticipated ramifications in particular circumstances, or because the simulation did not include enough rules to accurately represent how people behave in these situations.
In Blood & Laurels, we did all sorts of things to prevent the surprising surprises: NPCs don’t ever interact offstage, only in the player’s presence, so there’s no invisible shuffling of NPC allegiances or knowledge states; NPCs don’t lie to the player except in clearly defined and authored situations; most scenes take place in a single location and NPCs are not allowed to wander off; each scene is marked to indicate whether the extreme simulation outcomes are allowable. (So for instance, you can set a scene to indicate that no one may leave the room, have sex, or die during that scene.)