Argument Box Rehaf AlJammaz, 1 Yasheng She,2 Michael Mateas3 1 raljamma@ucsc.edu 2 yashe@ucsc.edu 3 michaelm@soe.ucsc.edu University of California, Santa Cruz Abstract create characters with better behavior understandability, that are less predictable, and have more sophisticated awareness Games currently feature simple models of morality and moral of different social situations. All of which are dimensions of reasoning. These typically take the form of a simple binary opposition between good and evil, in which game actions believability (Gomes et al. 2013). How a character reasons are sorted into these categories, and reputation systems that about these moral situations can enhance the illusion of be- track the reputation of players (and sometimes NPCs) with in- lievability; one aspect of reasoning involves belief modeling. game factions based on the actions taken. This paper presents When referring to belief modeling in games, we usually a more sophisticated model of moral reasoning based on associate it with facts and theories of mind. For example, a Lakoff’s metaphorical family models: the strict father and given character might believe that an NPC has brown eyes nurturant parent morality. This model utilizes more rapid, or another NPC likes them. Research games employ this fac- surface level categorization of different situations in moral tual sense of belief modeling, such as characters believing in categories, and deeper reasoning that characterizes situations in terms of their relationship to moral virtues as determined false information in (Ryan et al. 2015) and (Guimaraes, San- by the metaphor value system. This model is used in the ex- tos, and Jhala 2017). Beliefs under this category have also perimental prototype Argument Box (AB), a social argument been used alongside social systems (Evans and Short 2013; simulator where the player argues with clients visiting the Morais, Dias, and Santos 2019), creating unique stories. shop. Arguments in AB center around the moral virtues and They have even been used as game mechanics, such as the vices of simulated characters in the social simulation Talk Of player influencing NPC actions by directly ’injecting’ false the Town. This paper presents the current architecture, dis- beliefs in MKULTRA (Horswill 2015). In this paper, we re- cussing its technical details. fer to a broader definition of belief modeling that includes an NPC’s moral beliefs and values. We believe there is a need to create a more sophisticated Introduction moral reasoning system. This paper presents our current In many faction and reputation-based games, morality is of- progress in Argument Box: a game prototype that employs ten portrayed as a simple binary value of good and evil. moral reasoning for NPCs in terms of both surface and deep The reputation of the player character (and sometimes the values. Non-Player Character (NPC)) is represented as a single- Our prototype adds a layer of moral reasoning by incorpo- dimensional scale where players or NPCs land on a spectrum rating Lakoff’s metaphorical family models: the strict father based on their deeds and the morality points they accumu- and nurturant parent morality into our agent’s belief systems. late. In the current prototype, we focus on the Strict Father Model Unfortunately, NPCs in these types of games tend to em- (SFM), which, as the name suggests, (Lakoff 2010) employs ploy surface-level judgments about the characters they en- a metaphorical strict father figure as the head of the house- counter. For instance, games such as World of Warcraft and hold, guiding our agent’s actions. The SFM (Lakoff 2010) Neverwinter Nights (Blizzard 2004; ObsidianEntertainment believes that “the world is a dangerous place.” Therefore the 2002) employ NPCs that attack opposing factions or views “parent” needs to protect the children, become authoritative without any reasoning from the characters involved, usu- and teach them “right” from “wrong,” reward them if they do ally based on high-level decisions such as the character be- good, and punish them if they do wrong. Each family sys- longing to an opposing faction. These simplistic moral deci- tem employs a number of metaphors or values they believe sions and binary systems can affect a character’s believabil- in; for example, the SFM values strength. Anything that re- ity, making them feel mechanical, stereotyped, and less life- duces strength is seen as weak and immoral. like. By developing comprehensive moral models, we can Our model classifies different situations according to each Copyright © 2021 for this paper by its authors. Use permitted character’s held moral beliefs at a surface level. Surface- under Creative Commons License Attribution 4.0 International level reasoning is used to initially describe an NPC’s stance (CC BY 4.0). on an issue. We also incorporated deeper reasoning capabil- moral decision-making is usually tied to the player’s ac- ities that are referenced when an NPC strongly cares for a tions. Games such as The Witcher 3, Undertale, and those by given topic, granting an NPC the ability to fight for a pas- Telltale Games (CDProjekt 2015; Fox 2015; TellTaleGames sionately held belief when the player disagrees with them. 2004) place the player in high stakes-moral situations where Our deeper model allows for deeper reasoning that defines their action affects other NPCs and the story’s outcome; this NPC situations related to moral values as determined by the is usually limited to the player character, without modeling father model metaphor. NPC rationalization in moral dilemmas. Other games that do model NPC moral decision-making do so at somewhat Related Work superficial levels. For instance, NPCs in World of Warcraft (Blizzard 2004) and Farcry (LCGEntertainment 2004) use In this section, we review related work in terms of moral- high-level faction judgments, such as attacking NPCs in op- ity, reasoning, and social simulation as it relates to various posing factions on sight. elements of our project. In academia, morality systems in games often target moral theories or thought experiments. Works such as Togelius Social Simulation and Belief-Based Reasoning (Togelius 2011) and Nelson (Nelson 2012) examine Kant’s Several research games incorporate social simulation categorical imperative, which states, “Act only according to and reasoning in their system’s designs. Work such as that maxim by which you can at the same time will that PromWeek (McCoy et al. 2012) and successor CiF-based it should become a universal law.” Togelius’s prototype in- systems (Morais, Dias, and Santos 2019; Guimaraes, San- volves a procedural system that defines a new maximum rule tos, and Jhala 2017) utilize social physics as the main me- each time an event triggers, following the categorical imper- chanic of their design, influencing player conversations and ative principle. Nelson’s prototype started with game rules NPC interactions through social elements such as an NPC’s and allowed the player to break them, with the game adding relationship with other NPCs. new rules making the rule breaking universal. According to Unsurprisingly, social simulation and belief modeling are Nelson (Nelson 2012), this quickly led to havoc. Togelius ar- closely related to one another. Cif-CK (Guimaraes, Santos, rived at a similar conclusion noting the difficulty of design- and Jhala 2017) adapted the CiF architecture (McCoy et ing meaningful gameplay around universalization of player al. 2010) by incorporating belief modeling alongside their action. social networks, making it possible for characters to be- Other authors have built systems based on political di- lieve in false information. Comme il Faut - Exiles (CIF-EX) vides as well as belief modeling; Azad and Martens’s sys- (Morais, Dias, and Santos 2019) extended belief modeling tem, Lyra (Azad and Martens 2019; 2018) simulates char- beyond the interacting character, where other NPCs can in- acter interactions and conversations in politically charged fer relationships between characters in the world. groups. What is unique about Azad et al.’s system is their In Talk of the Town (TotT) (Ryan et al. 2015), a historical character’s ability to learn new biases and introduce it to town simulator, incorporates characters that can forget infor- their knowledge bases. mation, misremember and lie. Beliefs in TotT are affected Lakoff’s Moral Politics Lakoff is well known for by a character’s social network, what characters remember, his work on metaphor theory, where he identifies deep and the strength of the information supporting a given belief. metaphors, such as Life is a Journey or Love is War, that One interesting application of TotT is Bad News (Samuel et structure human cognition and the use of language (Lakoff al. 2016), where players converse with real actors portray- and Johnson 2008). In more recent work, Lakoff has pro- ing different NPCs. The actor acts on instructions from a de- posed two governing metaphors underlying political and veloper behind the scenes. We incorporate TotT’s generated moral reasoning. These two metaphors, the Strict Father and characters as topics of conversation in Argument Box. Nurturant Parent, govern the conservative and liberal (re- Other work, such as MKULTRA (Horswill 2015), fo- spectively) world views regarding the individual and their cused on belief modeling, where beliefs are ’injected’ into relationship to society. Our work on moral reasoning for an NPC’s knowledge base, allowing the player to manip- NPCs described in this paper builds on this work. ulate the NPC directly. In Versu’s (Evans and Short 2013) model of belief, agents can share public views of the world Lakoff’s (Lakoff 2010) strict father model employs a and have specific instances of individual false beliefs creat- metaphorical strict father as the head of a figurative house- ing unique gameplay experiences. hold. The strict father is seen as an authoritative figure that values strength, discipline and believes in rewards and pun- While social networks do not directly influence this ishments. The strict father views morality in terms of one’s project, we use characters that incorporate social relation- ability to abide by certain values such as strength, moral ships and elements in their design. We essentially used Talk order (following the natural order of things), and Moral of the Town’s (Ryan et al. 2015) exported characters as con- Boundaries (i.e., deviating from the norm is wrong). Any- versation topics between the player and the NPC. thing that reduces these values is seen as immoral. Lakoff illustrates an example where the act of purchasing illegal Moral Reasoning drugs is seen as immoral, as it emerges from low self-control Morality in Games There are many games that involve (i.e. low strength) according to SFM. morality and moral decision-making. In commercial games, On the other hand, the Nurturant Parent believes that members can grow as a result of nurturance and care. It en- to seed conversational topics by our CNPCs. These BNPCs courages self-reliance by caring for others. Unlike the SFM, are generated from TotT. TotT, as a reminder, is a historical the figurative parents in the NPM value their member’s opin- town simulator that generates characters, including various ions; the model employs respect between its members rather elements such as character relationships, locations, and jobs. than dreading punishments or expecting rewards. Like the We import the data from TotT in JSON format. We SFM, the NPM has metaphors and values it believes in, such then search for patterns, such as combinations of at- as Morality as Fair Distribution, Morality as Social Nur- tributes or social connections on BNPCs or temporal se- turance (e.g., strengthening social relationships with others quences undergone by BNPCs, in a process similar to story and mending those relationships), and Morality as Nurtu- sifting (Kreminski, Dickinson, and Wardrip-Fruin 2019; rance (i.e., being regularly empathetic). Max Kreminski 2021). BNPCs are assigned tags based on In our current prototype, we focus on the values employed the patterns that matched. We then filter the BNPCs by by the Strict Father Model (SFM). The deep values held by thresholding the number of tags (only BNPCs with enough this model include Moral Strength, Moral Boundaries, and tags are potential topics of conversation) and place the fil- Moral Wholeness. We will further describe how we incor- tered list of BNPCs into a priority queue based on the num- porated these metaphors into our system in a later section. ber of tags found and the quality of the tags. Tags that involve multi-character patterns are weighted more highly The prototype and Gameplay than tags resulting purely from within-character patterns. We provide examples of the sifting patterns we use in a later sec- In a classic Monty Python skit called ’Argument Clinic’ tion. (Monty Python ), a man approaches a clinician asking to Once we have our list of BNPCs, the CNPC brings up ap- buy an argument; he then proceeds to argue with the clini- propriate dialogue based on the tags present on the BNPCs. cian about arguments, debating if their argument is an argu- The starting dialogue is unbiased and simply expands the ment! Likewise, our prototype features an argument simu- tag into a textual utterance. For example, the tag familyPer- lator game, where NPCs walk into a shop called Argument son gets translated into “Have you heard that X has a large Box (AB) to procure arguments with the player character, family?” focusing on arguments about moral behavior. Each pattern in the system is further mapped to the The prototype described in this paper starts when an NPC CNPC’s surface values. These surface values denote how a initiates conversations about characters living in their town. CNPC generally feels about a set of tags; there is a many-to- They give the player a piece of gossip about a particular many mapping of tags to surface values which is described character they heard about, including their opinion and eval- in more detail below. The CNPC holds each surface value uation of that character’s actions. The player can agree, dis- with a strength of high, medium or low, indicating how pas- agree or vouch for the talked-about character by arguing for sionately the CNPC believes in that value. particular stances. Depending on the player’s response and As long as the player agrees with the judgements of the how passionately the NPC feels about the current argument, CNPC, the conversation will stay at the level of surface val- the NPC can reference their deep-seated beliefs as moral ar- ues. However, when the player disagrees with the CNPC guments, which for the presented prototype are based on on a surface value they hold strongly, the CNPC will per- Lakoff’s strict father model. In the following sections, we form deeper reasoning using the strict father model to back will discuss the system in greater detail. First, we examine up their claim. The conversational options provided to the the components of our system at a high level; we then look at player allow them to agree, disagree, bring up a specific dis- our NPC’s architecture, covering its creation, design, and be- cussion, or change the conversation topic entirely if they de- lief modeling. Lastly, we go through two examples covering sire. Figure 1 illustrates the general components of the sys- a typical argument in our simulation. We note that the char- tem. The following sections will describe how each compo- acters in Argument Box are modeled as polyhedra. Thus, in nent works in greater detail. the example dialog below, there are references to “shapes” which in our world is the same thing as saying “people.” Modeling BNPCs High Level Overview BNPCs are generated using the TotT simulation. Our sifting patterns focus on topics we can have moral debates about, Our current prototype features a character coming into a and so generally exclude details such as locations of homes clinic-like shop called Arg Box to argue with the player and businesses, street names and physical character descrip- character about the latest gossip in town. These charac- tions. We currently use 58 patterns to assign tags to BNPCs ters are called conversational NPCs (CNPC). The conver- as follows: sational NPCs (CNPC) are the main NPCs the player con- verses within the Argument Box. • Patterns that directly map a single TotT raw attribute. Ex- A single game loop involves the CNPC conversing about amples include isWealthy, departed, and familyPerson, in- one of the town’s characters, stating what they heard, and dicating if a character has wealth, left town, or has a fam- saying how they feel about it. The conversed-about charac- ily, respectively. ters are what we refer to in this paper as the background • Patterns that are created by combining different TotT raw NPCs (BNPC). The BNPCs include a separate list of NPCs attributes. For example, a retired character that is 45 years that the player character never interacts with; they are used old has the pattern retiredYoung assigned. how they feel about this tag by relating it to their surface values. There are currently 28 surface values defined in our sys- tem, of which a subset will be held (with varying strengths of low, medium or high) by a CNPC. Examples include Love- IsForFools, LoveAboveAllElse, FamilyPerson, and Shape- sArenothingIfNotSocial. BNPC tags map in a many-to-many way with surface values. The tags that map to a given sur- face value are called core tags of that surface value. The core tags have a many-to-many relationship with the surface val- ues. For example, the tag willActOnLove is a core tag of both the surface values BeTrueToYourHeart and LoveIsForFools. In the case that a core tag maps to two or more mutually held surface values, this provides some non-determinism on how Figure 1: Overview of the system, featuring the components the CNPC will comment on the presence of this tag, depend- of the player, BNPC and CNPC architectures ing on which surface value is taken as being activated. When a CNPC is instantiated, the system randomly as- signs the surface values and their accompanying strength. • Patterns that are based on TotT character jobs; these pat- Some surface values are mutually exclusive, so can not be terns are used to make controversial assumptions about simultaneously held with high strength. For example, if our characters, based on the character’s career and its poten- CNPC holds LoveIsForFools with a high rating, it cannot tial effect on others. For example, any TotT character with hold BeTrueToYourHeart with any strength other than low. the job miner or cooper is assigned the tag polluterRole. Additionally, the mapping from core tags to surface val- This is later used for conversations related to the environ- ues is used for the conversational options presented to the ment. player. This allows the player to bring up BNPC characteris- • Patterns that focus on relationships with other generated tics during the conversation that explicitly agree or disagree characters. such as love triangles and backstabbing. Ex- with the CNPC at the level of surface values. amples include friendWithBestFriendsEnemy and InLove- The surface values are used for immediate value-laden re- WithSpouseOfFriend. actions during the conversation. As long as the player and CNPC agree with each other, the conversation can stay at • Patterns that combine lower-level tags into higher-level the surface value level. However, when the player’s judge- ones. For example, if a BNPC has the tags adultButNot- ments disagree with the CNPC, the system switches to rea- working and IsWealthy (from lower-level patterns) this re- soning about deep values as determined by the Strict Father sults in the higher-level tag notWorkingAndRich. Model (SFM). Implementing a version of the Nurturant Par- • The last category of patterns combines tags found by ent Model (NPM) is ongoing work. Switching to this deeper other patterns with the Five Favor Personality traits (Gold- model allows the CNPC to marshal arguments by bringing berg 1990; Ryan et al. 2015) that TotT assigns to char- up characteristics that relate to more deeply held values. acters. For example, the pattern hasAlotOfEnemies com- This prevents the conversation from immediately degener- bined with the personality factor of high Agreeableness ating into repeated assertions (e.g. “Yes it is! Not it isn’t! results in the tag tooTrustingOfEnemies. Yes it is! Not it isn’t”) at the surface level. As mentioned above, once all pattern matching is com- Deep Values: Strict Father Model pleted and the tags have been assigned, BNPCs are placed in a priority queue based on the number and quality of tags. We define six deep values drawn from Lakoff’s book How By starting the conversation about characters at the front of Liberals and Conservatives think (Lakoff 2010) to specify the queue, this ensures that the CNPC will have a good num- the SFM: ber of debatable topics to argue about with the player. • Moral Boundaries warns about the danger of deviating from the norm. Characters that deviate from the norm are Modeling CNPCs seen as immoral by a character holding the SFM. Surface Values • Self Interest sees seeking one’s self-interest as moral and The CNPC starts the conversation by picking the BNPC at interfering with one’s self-interest as immoral. the front of the queue to talk about, doing this until it has • Moral Wholeness is concerned with unity and conformity exhausted all the BNPCs tags or the player chooses another among characters. character to talk about. Once a tag has been chosen, the CNPC starts the conversa- • Moral Essence evaluates a character’s past actions as in- tion by commenting on the tag in a value neutral manner. For dicators for their future actions, making the assumption example, if our BNPC named Mike had the tag familyPer- that past actions are the result of a character’s “essence”. son selected by the system, the CNPC states, “Oh, have you • Strength values a character’s ability to act in or handle heard that Mike has a big family?” The CNPC then states difficult or sensitive situations. Low strength is seen as Figure 2: Surface value conversational loop, player agrees moral weakness. stance. • Moral Order accounts for traditional hierarchical power relationships, such as rich characters viewed as morally Example: Surface Value Conversation superior to poor ones. Our first scenario will illustrate a typical conversational loop As we mentioned earlier, each surface value includes a showing what happens when the player agrees with our set of core tags. Our SFM deep values provide a mapping CNPC on a medium-rated surface value. for each core tag to high morality or low morality. High At the start of the loop, the CNPC picks a BNPC to indicates that the core tag exemplifies the deep value; low converse about, in this case Caroline Milliem. Caroline has indicates it violates the deep value. For example, the sur- three tags available, inLoveWithSpouseOfFriend, familyPer- face value BeTrueToYourHeart and one of its core tags, son and willActOnLove. inLoveWithSpouseOfFriend is evaluated as low by the Moral The CNPC selects Caroline’s first pattern, inLoveWith- Boundaries deep value. If our CNPC supports the surface SpouseOfFriend, and opens the conversation by gossiping: value BeTrueToYourHeart and is involved in a argument “Oh, Have you heard! Caroline Milliem is In love with their with the player over this value, it would not make use of friend’s spouse!” Under the hood, the pattern was mapped the BNPC tag inLoveWithSpouseOfFriend as an argument to the surface value LoveIsForFools as ”inLoveWithSpouse- for their claim, as this violates a deep value. Our upcoming OfFriend is one of its core tags. We note, this assignment examples will further clarify these situations. happens at random; it could have mapped to other surface Furthermore, the SFM can check if a CNPC is searching values as long as the pattern belonged to that core set of the for pro or contra argument to back up a claim about a given selected surface value. surface value. Generally, if a CNPC is supporting an argu- It so happened that our CNPC is relatively indifferent to ment for a given surface value, the SFM is looking for core this value, holding it with medium strength. They care about tags that evaluate as high against the deep values, possibly the value but not so much that they would raise a fuss if the switching the argument to another surface value if no such player disagreed with their opinion (represented as a “dis- tags can be found for the current surface value (this is an agree” button). The CNPC then presents its opinion based on example of deflecting to another topic to continue to argue this surface value, stating “Well, I think that shapes should for the moral virtue or vice of the BNPC if the CNPC can never let their emotions cloud their judgment.” no longer argue for the current surface value). Alternatively, The player then has the option to agree, disagree or bring the SFM can look for core tags that evaluate as low against a up another topic or BNPC as clickable options. In our sce- deep value to provide a cautionary tale in the argument. For nario, the player agrees: “I couldn’t agree with you more.” example, if the CNPC holds the surface value LoveIsFor- The CNPC, satisfied with the result, says, “Yeah, it’s good Fools, the SFM can look for a core tag, such as inLoveWith- that you see it my way!”. We note that this scenario presents SpouceOfFriend, that evaluate as low against a deep value a medium stance on a surface value; if the player disagrees, (e.g. Moral Boundaries), to make a negative statement about the CNPC simply notes its distaste for the player’s choice. the CNPCs behavior, and thus support their LoveIsForFools The system then checks if the other tags of the current Figure 3: Deep value conversation loop - player disagrees with high surface value BNPC have been explored; if they haven’t been explored, We note that the text is written in a way that references the the system proceeds with the same loop but with the newly specific tag, illustrates the deep value, and provides reason- appointed tag familyPerson. Otherwise, the system moves ing as to why the surface value LoveIsForFools should be on to the next BNPC in the queue. held. The player can then agree or disagree with the pattern, choose another BNPC pattern to bring up, passing it as an A Deep rooted conversation - Using the SFM argument for the current surface value. Let us take the previous example but assume our CNPC ac- If the CNPC is presented with a tag that the SFM maps tually cares deeply about the surface value LoveIsForFools. in a way contradictory to the surface value argument being The CNPC conveys this as, “Well, I think that shapes should made, it then searches through the BNPC’s available tags for never let their emotions cloud their judgment AND doing an alternative argument. If any of the remaining tags are core what’s right is always better than doing what feels right.” As tags of the surface value and the SFM mapping supports the a note on authoring, we use AND and BUT to emphasize argument, it presents it as a backup argument. Otherwise, the the CNPC’s high or low stances by extending the same sen- CNPC mimics a person backed into a corner, randomly fir- tence with a modifier that enhances how strongly or weakly ing off defenses based on the tags found. This results in more they feel about a given topic. This helps minimize the com- generic arguments, for example, stating “But that BNPC has binatorial amount of dialog we have to write and signals the a family!” in response to the situation where the BNPC is in underlying model more strongly to the player. love with the spouse of a friend. In this example the player disagrees with the CNPC, se- lecting the dialog option: “Are you kidding?? Love is the best thing ever.” Again, the tone mimics that of the CNPC. Current Limitations and Future Work The CNPC then responds with, “You’re joking! Love is for fools.” It consults the underlying model, in this case the This paper presents our current work on Argument Box, SFM, to back up why love is indeed for fools. a game prototype incorporating a moral reasoning system The currently selected tag inLoveWithSpouseOfFriend is that operates on both surface and deep values (beliefs). validated as a core tag of LoveIsForFools, and determined The model is inspired by Lakoff’s work on family-based to score low against the deep value Moral Boundaries. Thus metaphors in moral reasoning. As future work we are im- the SFM validates the surface value stance, and allows the plementing the Nurturant Parent Model, refining the con- CNPC to make an argument based on deep values: “Love versation loop so that under appropriate (but challenging) can be immoral. Honestly, they have no boundaries. This circumstances the player can change the CNPCs mind, and shape went after their friend’s partner; that’s just wrong.” playtesting the game. References ObsidianEntertainment, B. 2002. Neverwinter nights. [CD- Azad, S., and Martens, C. 2018. Addressing the elephant in ROM]. the room: Opinionated virtual characters. In AIIDE Work- Ryan, J.; Summerville, A.; Mateas, M.; and Wardrip-Fruin, shops. N. 2015. Toward characters who observe, tell, misremem- Azad, S., and Martens, C. 2019. Lyra: Simulating believable ber, and lie. In Proceedings of the AAAI Conference on opinionated virtual characters. In Proceedings of the AAAI Artificial Intelligence and Interactive Digital Entertainment, conference on artificial intelligence and interactive digital volume 11. entertainment, volume 15, 108–115. Samuel, B.; Ryan, J.; Summerville, A. J.; Mateas, M.; and Blizzard. 2004. World of warcraft. CD-ROM. Wardrip-Fruin, N. 2016. Bad news: An experiment in com- CDProjekt. 2015. The witcher 3: Wild hunt. Xbox One. putationally assisted performance. In International Confer- ence on Interactive Digital Storytelling, 108–120. Springer. Evans, R., and Short, E. 2013. Versu—a simulationist sto- rytelling system. IEEE Transactions on Computational In- TellTaleGames. 2004. Telltalegames. PC. telligence and AI in Games 6(2):113–130. Togelius, J. 2011. A procedural critique of deontological Fox, T. 2015. Undertale. PC. reasoning. In DiGRA Conference. Goldberg, L. R. 1990. An alternative” description of person- ality”: the big-five factor structure. Journal of personality and social psychology 59(6):1216. Gomes, P.; Paiva, A.; Martinho, C.; and Jhala, A. 2013. Met- rics for character believability in interactive narrative. In International conference on interactive digital storytelling, 223–228. Springer. Guimaraes, M.; Santos, P.; and Jhala, A. 2017. Cif-ck: An architecture for social npcs in commercial games. In 2017 IEEE Conference on Computational Intelligence and Games (CIG), 126–133. IEEE. Horswill, I. D. 2015. Mkultra. In Eleventh Artificial Intelli- gence and Interactive Digital Entertainment Conference. Kreminski, M.; Dickinson, M.; and Wardrip-Fruin, N. 2019. Felt: a simple story sifter. In International Conference on Interactive Digital Storytelling, 267–281. Springer. Lakoff, G., and Johnson, M. 2008. Metaphors we live by. University of Chicago press. Lakoff, G. 2010. Moral politics: How liberals and conser- vatives think. University of Chicago Press. LCGEntertainment. 2004. Far cry primal. PC. Max Kreminski, Melanie Dickinson, M. M. 2021. Winnow: A domain-specific language for incremental story sifting. In Conference on Artificial Intelligence and Interactive Digital Entertainment. McCoy, J.; Treanor, M.; Samuel, B.; Tearse, B.; Mateas, M.; and Wardrip-Fruin, N. 2010. Comme il faut 2: A fully real- ized model for socially-oriented gameplay. In Proceedings of the Intelligent Narrative Technologies III Workshop, 1–8. McCoy, J.; Treanor, M.; Samuel, B.; Reed, A. A.; Wardrip- Fruin, N.; and Mateas, M. 2012. Prom week. In Proceedings of the International Conference on the Foundations of Digi- tal Games, 235–237. Monty Python, u. c. Argument - monty python. Morais, L.; Dias, J.; and Santos, P. A. 2019. From caveman to gentleman: a cif-based social interaction model applied to conan exiles. In Proceedings of the 14th International Conference on the Foundations of Digital Games, 1–11. Nelson, M. 2012. Prototyping kant-inspired reflexive game mechanics. In Proceedings of the 2012 Workshop on Re- search Prototyping in Games.