# Curating Simulated Storyworlds — Chapter 4: The Pain of Emergent Narrative

> NOTE: THIS IS A MARKDOWN CONVERSION INTENDED FOR LLM CONSUMPTION.
> IF YOU ARE A HUMAN BEING, READ THE PDF INSTEAD:
> https://jamesryan.computer/papers/2018_ryan_curating_simulated_storyworlds.pdf
>
> The PDF is the canonical document, typeset with every figure and table. This file
> is a plain-text conversion of the LaTeX source, published HERE so that software agents
> can read the thesis without parsing a PDF. It is lossy and is not meant for humans.
>
> This file is chapter 4 of 14 only: The Pain of Emergent Narrative (~41,000 words).
> See below for links to markdown conversions of the individual chapters.
>
> **Structure.** Sections are `##`/`###`/`####` headings. Cross-references are resolved to
> their numbers ("Section 3.1.1", "Chapter 5"), so a heading can be searched for directly.
> Citations are inline bracketed keys such as `[ryan2017grimes]`; every key used in the text
> has an entry in the bibliography, formatted as `- **[key]** Author. Title. Venue, Year.`
> Footnotes appear inline as `[Note: ...]`.
>
> **Conversion limits.** Figures and most complex tables appear as `[Figure omitted]` or
> `[Table omitted]`. Code fences carry no language tag. A few diacritics in author names
> were simplified from the Latin-1 source.
>
> **All files.**
>
> Index (this list, plus the abstract; small): https://jamesryan.computer/thesis/index.md
> Whole thesis in one file (large, ~1.6 MB): https://jamesryan.computer/thesis/full.md
>
> By chapter:
> - 01. Introduction (~4,600 words, ~30 KB) — https://jamesryan.computer/thesis/01-introduction.md
> - 02. Defining Emergent Narrative (~6,900 words, ~47 KB) — https://jamesryan.computer/thesis/02-defining-emergent-narrative.md
> - 03. The Pleasure of Emergent Narrative (~21,000 words, ~135 KB) — https://jamesryan.computer/thesis/03-the-pleasure-of-emergent-narrative.md
> - 04. The Pain of Emergent Narrative (~41,000 words, ~268 KB) — https://jamesryan.computer/thesis/04-the-pain-of-emergent-narrative.md
> - 05. Refining Emergent Narrative (~11,000 words, ~73 KB) — https://jamesryan.computer/thesis/05-refining-emergent-narrative.md
> - 06. James Ryan Generator (~2,800 words, ~16 KB) — https://jamesryan.computer/thesis/06-james-ryan-generator.md
> - 07. World (~20,000 words, ~129 KB) — https://jamesryan.computer/thesis/07-world.md
> - 08. Case Study: Diol/Diel/Dial (~9,300 words, ~59 KB) — https://jamesryan.computer/thesis/08-case-study-diol-diel-dial.md
> - 09. Talk of the Town (~25,000 words, ~162 KB) — https://jamesryan.computer/thesis/09-talk-of-the-town.md
> - 10. Case Study: Bad News (~22,000 words, ~132 KB) — https://jamesryan.computer/thesis/10-case-study-bad-news.md
> - 11. Hennepin (~32,000 words, ~209 KB) — https://jamesryan.computer/thesis/11-hennepin.md
> - 12. Looking Ahead: Sheldon County (~11,000 words, ~75 KB) — https://jamesryan.computer/thesis/12-looking-ahead-sheldon-county.md
> - 13. Conclusion (~2,700 words, ~17 KB) — https://jamesryan.computer/thesis/13-conclusion.md
> - 14. Bibliography (~29,000 words, ~233 KB) — https://jamesryan.computer/thesis/14-bibliography.md

The essential critique of emergent narrative is that stories produced by the method are often bad, or moreover, are not even stories at all. In this chapter, I unpack this critique to discuss a series of painful shortcomings that have plagued emergent narrative since its inception in the middle of the last century. As I will argue in the next chapter, none of these points represent inherent limitations to the approach, as some have suggested, but rather shortcomings in a number of its notable works. As in the last chapter, I will refrain from discussing my own practice here, but I will note that in Part II, I critique my projects (sometimes brutally) with respect to the issues that I identify here. I hope that this chapter provides insights that enable not only improved crafting of emergent narrative, but also refined interpretation of works in the area—more specifically, particular misunderstood efforts that are of historical importance. Finally, while my primary aim here is to identify a series of pitfalls that have worked to hamper the form, in each case I will outline potential solutions, many of which have quietly been implemented in existing works. As such, this chapter also loosely suggests a *poetics of emergent narrative*.

## Simulation Pains

Since at least Natalie Dehn's 1981 paper “Story Generation After *Tale-Spin*” [dehn1981story], practitioners have identified a fundamental pitfall to the simulationist approach taken in emergent narrative: when characters are given free reign, the simple result of running a given simulation instance may not be a good story. As Dehn put it, “Setting up an initial story world such that things work out in an interesting way is, at the very least, tricky business” [dehn1981story, p. 16]. Indeed, even in her seminal first paper on the topic, Ruth Aylett noted, “one of the risks of emergent narrative is that it may not emerge—the unpredictability that makes it interesting also makes it in some sense fragile” [aylett1999narrative, p. 85]. Thus, works of emergent narrative may be doomed before the narrative ever makes it to the surface.

In this section, I identify and discuss a series of issues that pertain to the crafting of the underlying simulation that models and evolves the storyworld out of which narrative may emerge. These issues may be viewed as design pitfalls that simulation authors should take into consideration when developing their own systems. Some of these problems are germane to the broader area of story generation—or, even more generally, to procedural generation and computer simulation—while others seem to be endemic to emergent narrative.

### Boringness

A fundamental pitfall in emergent narrative pertains to the level of interestingness inherent in the simulation domain—that is, some simulations are too *boring*. In this section, I will explore this idea by touching upon scholarly investigation of the notions of *narrative criteria*, *story interestingness* and *tellability*.

#### Criteria for Narrative

In the most extreme cases, a simulation may be so lacking in the intrigue that we demand of stories that it cannot generate any good stories, or even anything that could be called narrative at all. While this critique has been levied against systems like Jim Meehan's *Tale-Spin* [meehan1976metanovel], an actual extreme case would be something like the famous *blocks world* of Terry Winograd's *SHRDLU* system [winograd1972understanding]. Many projects have used this microworld as a testbed for planning [slaney2001blocks], and in such projects the trace of a path to a goal state works just like story generation works in *Tale-Spin*—the generated output is a *simulation trace*—but the resulting traces would never be called stories. This is intuitive, but still curious, because such a project satisfies the basic *narrative criteria* implied by Meehan's design of *Tale-Spin*: a story occurs when an entity has a goal and executes a plan that terminates in either the goal state or a failure state. [Note: Several researchers around this time identified problem solving as a critical feature of narrative [rumelhart1976understanding,black1980story]. Robert Wilensky noted that obstacles to solving such problems are equally important [wilensky1982points, p. 354]. The importance of an obstacle is taken for granted in the remainder of this section.]

[Figure omitted]

So what is missing? The most glaring omission is characters—let us say then that the blocks world and other microworlds without characters (a lunar base [woods1972lunar], a fueling station [heidorn1971natural], a power circuit [brown1974sophie], etc.) are probably not capable of producing narrative. Let us then consider one that does feature characters: the *clowns world* of Robert F. Simmons, which concerns a clown who can stand on his head, do tricks with a pole, and sail to a lighthouse on a boat [simmons1975clowns]; Figure 4.1 shows the system in action. Is this storyworld capable of producing narrative? The issue is murkier, but conveniently in this case we can say it cannot because the clown does not pursue goals beyond executing individual tricks that are specified by a human user. A later microworld called *Simon City* reintroduces the confusion, however: it is the home of a simulated taxi driver named *Elmer*, who executes plans in order to drive around the city [mecalla1979execution]. [Note: Beautifully, the city is named for AI pioneer Herb Simon, and its streets are named for other well-known figures in the field: Winograd Way, Schank Strip, the Minsky–McCarthy Freeway. I find the early microworlds to be quite evocative when they are viewed as aesthetic objects—a kind of simulationist art brut, in line with my argument in Section 3.1.4.] This is a virtual character who has goals and executes plans in order to overcome the obstacles to achieving those goals—there is even a notion of other characters who stand in the way: modeled drivers that may inhibit *Elmer*'s driving. Is a trace through *this* simulation a narrative? It seems like it could be—and it is certainly closer to being one than the blocks world and the clowns world—but certain features still seem to be missing.

Let us return to Meehan's conception of *Tale-Spin*, which he articulates, in the abstract to a conference paper on the system, in terms of a more extensive set of narrative criteria :

> Tale-Spin is a program that writes stories by using knowledge about problem solving, physical space, interpersonal relationships, character traits, bodily needs, story structure, and English. [meehan1977tale, p. 91]

 Leaving aside both ‘English' (since our purview here is an underlying simulation) and the teleological feature ‘story structure'—this is useless to us here, since we are trying to figure out what exactly that means—perhaps we can say that what is missing in *Simon City* is interpersonal relationships, character traits, and bodily needs. Here, then, is a computer-generated paragraph that has confounded me in my project of uncovering the forgotten early history of story generation [ryan2017grimes]:

> Mary was wrestling with a bottle at the liquor bar before John helped her with it. John saw Mary wresting with a bottle at the liquor bar. John went over to help Mary with a bottle before he drew a bottle cork. John and Mary together drank the champagne. [simmons1972generating, p. 903]

 The paragraph was generated by a project by Robert F. Simmons and Jonathan Slocum in 1972. Simmons also created the *clowns world*, mentioned above, and he was Sheldon Klein's collaborator on some of the earliest work in natural language generation [simmons1962toward,klein1963syntactic,klein1963computational,simmons1964indexing]. In fact, Klein's later work in story generation [klein1971program,klein1973automatic,klein1974modelling] evolved out of this collaboration, so we can say that those projects and the one at hand are cousins. While Simmons and Slocum were not interested in story generation, [Note: Later on, toward the end of 1970s, Simmons and his student Alfred Correira would produce *Telltale*, one of the major early story generators [simmons1979rule,correira1980computing].] this paragraph is reminscient of the prose of Klein's generated murder mysteries [klein1973automatic], discussed at length below. But while Klein's outputs are clearly stories, this paragraph does not seem to be as narrative. [Note: In a related example, Robert Wilensky argues that the famous Schankian “restaurant story” [schank1975scripts, p. 153] is in fact not a story at all [wilensky1982points, p. 346]. As I show momentarily, the example in question here is in fact more challenging to account for, since it would actually still meet Wilensky's revised criteria for a story (problems with obstacles, as discussed below).]

Why is this? Linking back to Meehan's story criteria, we can say that the paragraph pertains to problem solving (how to open the bottle), physical space (the bar), interpersonal relationships (John vis-à-vis Mary), character traits (John seems helpful), and arguably even bodily needs (perhaps they each went to the bar as a way to meet someone). So why is this not quite a story? It has all of Meehan's components, and it also has explicit causality, which in Section 4.1.6 I will argue is a necessary condition for narrative. Perhaps the framing of Simmons and Slocum is central—after all, they do not say it is a story. [Note: This issue of framing is very interesting to me, since it seems to be the heuristic with which scholars delineate the tradition of computational art from concurrent developments in other modes of computational work. Here, a prime example pertains to two of the earliest text generators, both of which emerged out of early American work in that area. The first of these, *Auto-Beantik*, was framed as an effort in computer poetry [worthy1961librascope,worthy1962new], and it has been remembered as a computational art project accordingly [newell1983pattern,hartman1996virtual,funkhouser2007prehistoric]. Contemporaneously with that project, Victor Yngve utilized essentially the same method (context-free grammar) to produce syntactically well-formed nonsense sentences, partly in an effort to test a theory of English syntax [yngve1961random]. Yngve's system has not been remembered as computer poetry, but instead it shows up in the (scant) histories of natural language generation, where intriguingly early efforts in computer poetry do not appear [mann1982text,mckeown1992text,kantrowitz1993bibliography,bateman1997sentence]. (Fascinatingly, however, one strangely enchanting sentence generated by Yngve's system would charm a number of poets: “What does she put four whistles beside heated rugs for?” [milic1970possible, pp. 4–5][plumb1964scholars][smith2001world, p. 129][vanhorne1964funny][preminger2014princeton, p. 43].) At the time, the systems were viewed as the two prominent examples of text generation [garvin1962computer, p. 387], but eventually they would be remembered according to how their creators respectively framed them. I am interested in viewing all scientific projects as artistic endeavors and all artistic endeavors as scientific projects—by doing this, we reveal a new series of evocative aesthetic objects and a more robust history of human inquiry. Indeed, in my digging I have found that most of the very earliest advances in natural language generation by computer—from Strachey's love-letter generator [strachey1954thinking,wardrip2011digital] to early computer poets both known [funkhouser2007prehistoric] and unknown (which I hope to report on soon)—were computational art projects. It is both exciting and disheartening to imagine how some of the early AI efforts could have instead been monumental mid-century artworks. For example, I think Robert Abelson's *Goldwater Machine* [abelson1965computer,abelson1973structure,wardrip2009expressive] instantiated a fundamentally new kind of representational artistic method. If it had instead been framed as an aesthetic contribution, I think it would be remembered today as a touchstone artwork, and perhaps even the progenitor of a distinct simulationist movement.] While this is probably a contributing factor, I think the answer is a different but equally simplistic one.

#### Story Interestingness

Some simulation domains are so boring that they seem to be fundamentally incapable of producing narrative, or at least anything that would be considered compelling narrative. So what makes a domain interesting? To be clear, by domain I refer generally to the stuff of the simulation: everything that it models, including setting, characters, attributes, actions, and so forth. To me, it seems that the most critical issue is the kinds of actions that are possible given what has been authored for the domain model. *Simon City*, for example, models a domain that I find inherently interesting (cities), but its sole action type (*Elmer* the taxi driver driving around) lacks intrigue.

Here, we can turn to the rather substantial body of work on *story interestingness*, which was a popular topic of research among cognitive scientists in the 1980s [schank1979interestingness,kintsch1980learning,wilensky1983story,britton1983makes], and one that continues to receive attention today through work by researchers such as my colleague Morteza Behrooz [behrooz2017modeling,behrooz2018cognitive]. A seminal paper in this area is Roger Schank's “Interestingness: Controlling Inferences” (1979), in which he posits three mechanisms that drive story interestingness [schank1979interestingness]. The first of these is an operator that he calls *unexpected events*: stories in which things happen unexpectedly for a given listener may be interesting to that listener. Next, he identifies another operator, *personal relatedness*, which works similarly: a story about oneself or someone one knows may be inherently interesting. Note, however, that Schank identifies these mechanisms not as being categorical in function, but as *operators* that take the single argument of a topic. Specifically, Schank identifies (as a third mechanism for interestingness) a set of topics that he calls *absolute interests*: danger, death, power, sex, money, destruction, chaos, romance, disease, “and many other concepts and issues of this type” [schank1979interestingness, p. 15]. As such, the larger idea is that stories may be interesting for pertaining to the topics of absolute interest, and they may be especially interesting if such topics are amplified by the operation of unexpectedness and or personal relatedness: “If someone you knew died from having sex for a lot of money, that would be very interesting by my rules” [schank1979interestingness, p. 16].

As Behrooz notes, Schank's idea of absolute interests was supported in articulations of later related concepts including Robert Wilensky's “human dramatic situations” [wilensky1982points,wilensky1983story], Peter Freebody and Richard C. Anderson's “generically important elements” [freebody1986serial, p. 32], and Behrooz's own “instinctive interests” [behrooz2018cognitive, p. 3]. Others have referred to this as *situational interest* [krapp1989interest,hidi1990interest,hidi1992situational]. Unpacking Schank's notion of unexpectedness, Wilensky notes that storytelling is predicated on generating expectations (in the mind of the recipient) as to what comes next [wilensky1982points,wilensky1983story], which makes story understanding a kind of problem solving [black1980story]. In this regard, Walter Kintsch identifies a specific pattern by which particular assumptions are evoked (or even elicited) in the mind of a listener as a clever set-up for violating those assumptions later on [kintsch1980learning]. George Mandler posited a similar theory a few years later [mandler1982structure], and other researchers have since corroborated this idea through experimental investigation [frick1992interestingness,campion2009contradictions]. Additionally, Bruce K. Britton and Asghar Iran-Nejad independently key in on the importance of the listener eventually arriving at a state of resolution with regard to the violation of her assumptions [britton1983makes,iran1987cognitive], through what Sung-il Kim identifies as an act of inference [kim1999causal]. Generally, researchers in this area are interested in the way that stories generate interest by stimulating cognitive procedures in the mind of a recipient—this is what Kintsch calls *cognitive interest* [kintsch1980learning, p. 88]. Notably, this phenomenon connects also to cognitivist narratology and narrative psychology, where narrative is considered as a mechanism for mental simulation [escalas2004imagine,zunshine2006we,oatley2011such,markman2012handbook,bormann2015immersed]. [Note: Intriguingly, inasmuch as the act of *crafting* a story is viewed as being based in a kind of mental simulation, the peculiar model of creative writing implicit in early story generators like *Saga II* [morse1960saga] and *Tale-Spin*—that an author comes up with a story by simulating a world in her mind (see Section 4.2.2)—is supported decades later by social science.]

As part of their ongoing project, Behrooz and his collaborators provide an additional set of useful interestingness factors: *common themes*, such as ‘being an underdog, growing up poor, being bullied in school”; idiosyncratic *topic interests*, examples of which might include “geography, sci-fi movies, fireworks”; a variant of Schank's personal relatedness that pertains more specifically to a kind of *reminiscence* about one's past; and the *implicit familiarity* of subject matter that evokes a visceral sense of familiarity or recollection in the recipient, such as a red rose [behrooz2018cognitive, p. 3]. Finally, Selmer Bringsjord and Dave Ferrucci make use of a notion similar to Behrooz's common themes, which enables an “engineering trick” that they utilize in their story generator *Brutus*:

> One of our tricks for ensuring (at least a degree of) interestingness is this: Formalize immemorial themes present in belletristic fiction, and have Brutus generate stories in accordance with this formalization. [bringsjord1999brutus, p. 6]

#### Tellability

Lastly, before connecting these ideas back to our discussion of simulation concerns, let us consider a body of work that concerns the more fundamental issue of whether a story is even *tellable* in the first place. This area originates in the work of sociolinguist William Labov (and collaborators), who identified the notion of a *reportable event* [labov1968study] as one whose reporting (in a narrative) is socially merited:

> telling a narrative requires a person to occupy more social space than in other conversational exchanges—to hold the floor longer, and the narrative must carry enough interest for the audience to justify this action. Otherwise, an implicit or explicit “So what?” is in order [labov1997narrative]

 Defined more concisely, “A reportable event is one which justifies the automatic reassignment of speaker role to the narrator” [labov1997some, p. 405], and indeed Harvey Sacks emphasizes that this process requires ongoing maintanance [sacks1995lectures]. Labov notes that certain kinds of events are always reportable—for instance, death and injury [labov1968study, p. 301]—an idea that corresponds to Schank's absolute interests, Wilensky's human dramatic situations, and other kindred notions. More broadly, a reportable event is one that is new (and newsworthy) or unexpected [sacks1995lectures,labov1997narrative,norrick2005dark].  Neal Norrick lists some additional sufficient criteria: “uniqueness and/or humor” and “the prospect of conarration” [norrick2005dark, p. 323]. [Note: Following Elinor Ochs and Lisa Capps [ochs2001living], Norrick defines a parabolic tellability, with boringness on one side and taboo (vulgarity, oversharing) on the other. Mary Louise Pratt, however, emphasizes the tellability of scandalous events [pratt1977toward, pp. 136–147], though such material may actually reside at the precipice of actual taboo.] Interestingly, Labov describes the construction of oral narrative as a causal backchaining from a “most reportable event” to the terminus of an event that does not require reporting:

> Before a narrative can be constructed, it must be pre-constructed by a cognitive process that begins with a decision that a given event is reportable. Pre-construction begins with this most reportable event and proceeds backwards in time to locate events that are linked causally each to the following one, a recursive process that ends with the location of the unreportable event—one that is not reportable in itself and needs no explanation. [labov2006narrative, p. 37] [Note: Later, I propose an operationalization of this process by which a system may automatically identify tellable event sequences that emerge from simulations.]

A more general term for Labov's ‘reportability' is *tellability*, which may be conceived as pertaining not strictly to the events of a story, but more broadly to its *stuff*. As Wolf Schmid notes, tellability may actually be predicated on the absence of expected events [schmid2010narratology, p. 13]. Indeed, for Marie-Laure Ryan, tellability is all about the suggestion of actions that do not occur (more precisely, that occur in plausible alternative possible worlds):

> Why is it that deception makes better tales than honesty, error than insight, broken promises than fulfilled obligations, temporary failure than continued success? Because all the above situations generate unrealized projections, and weave several parallel stories into the thread of the historical events. [ryan1986structure, p. 138]

 Moreover, as Ryan explains, and as others have as well [viktor1965art,labov1972language,prince1988disnarrated], many of storytelling's classic crosscultural tropes are rooted in such unrealized projections:

> Among the situations that involve virtual strings of events, we find themes of such common occurrence across cultures, periods, and genres as unsuccessful action, broken promises, violated interdictions, mistaken interpretation, and double, as well as single deception. [ryan1991possible, p. 158]

 Brilliantly, Ryan shows how these projections, diegetically embedded into the minds of characters, actually have a narrative structure (the poetics of which is itself a full-fledged narratology [ryan1986embedded, p. 328]). Ryan calls such cases *embedded narratives*: “story-like representation[s] produced in the mind of a character and reproduced in the mind of a reader” [ryan1986embedded, p. 320]. [Note: In a very interesting ongoing project, Ben Kybartas and collaborators are developing a videogame that is driven by an operationalization of Marie-Laure Ryan's possible-worlds narratology. The game is called *Subject and Subjectivity* [kybartas2017subject], and it is also showcases conversational interaction and text generation.]

David Elson argues that tellability is rooted ultimately in how a story modulates the basic needs of its characters through what he calls its *thematic content* [elson2012modeling, p. 45]. In Elson's *story intention graph*—by my estimation, the most deeply considered narrative schema in existence—character plans are represented as graphs whose sink nodes work as “affectively-charged termini” for those plans [elson2012modeling, p. 115]. By this configuration, story content is tellable when it *affects* a character, with regard to her basic needs. [Note: In his dissertation, Elson identifies twelve such basic needs (motivated by the work of Abraham Maslow [maslow1943theory] and Manfred Max-Neef [max1992development]): life, health, ego, wealth, love, leisure, membership, actualization, freedom, justice, enlightenment, and honor. Again, we find connection to Schank's categories of absolute interest, but also to Ryan's embedded narratives, inasmuch as character desires (and dreams and aspirations and fears) may pertain to such categories. It is worth noting, however, that content affecting such character needs will not automatically produce a good story, or a story at all. For instance, I believe that a fundamental problem with *Tale-Spin* is that its stories tend to be about uninteresting character goals, like a bird seeking a worm. While a bird's procurement of a worm will affect her health (recall Meehan's story criterion pertaining to bodily needs), that does not mean that a (good) story will necessarily obtain. To be clear, I do not mean to imply that Elson would claim as much—he would not—but it is worth noting nonetheless.] This is particularly useful, since most scholars of tellability aim to describe it as a sociolinguistic epiphenomenon of conversation, rather than to provide a framework for producing tellable stories. [Note: Generally, the work on interestingness and tellability is descriptive, not prescriptive, but the high-level insights are still invaluable for simulation crafting toward compelling emergent narrative.]

Finally, it is worth noting that tellability is ultimately established at the level of discourse, not in the reported event sequence. What this means is that good storytellers may render otherwise untellable events tellable, while bad storytellers may do the inversion. [Note: In her PhD thesis, titled *Unputdownable*, Beth Cardier writes about tellability from the standpoint of a narrative practitioner [cardier2012unputdownable].] Indeed, Jim Meehan has made this very point [meehan1982stories, p. 458]. Moreover, tellability is socially and culturally contextual, which means that a tellable event sequence may not always be tellable, even if the telling remains identical. For example, a riveting account of a strange dream will likely be better received by one's partner in the morning than in the middle of the night. When it comes to the prospect of automatic recognition of tellable events, the takeaway here is that the events cannot do all of the heavy lifting—ultimately, a horrible recounting of an interesting occurrence makes for a horrible story.

#### Simulating Toward Surface Intrigue

Let us return now to the specific concerns of simulation. While much of the knowledge derived from the scholarly work outlined in this section pertains to storytelling at the level of discourse, [Note: Accordingly, I will revisit these ideas later on in my discussion of the considerations surrounding curation in emergent narrative.] there are still plenty of ideas that may help to ensure that a simulation domain has inherent intrigue. As such, these ideas should be of use to practitioners who seek to simulate storyworlds that may produce tellable event sequences in the first place: ones that, when recounted in a generated story, will not elicit the dreaded “So what?” that is famous in the tellability literature [labov1972language,polanyi1979so,labov1997narrative,sacks1995lectures,labov2010should].

The most immediately pertinent of these ideas is Schank's notion of absolute interests, including topics such as danger, death, sex, money, destruction, chaos, power, and romance. Along such lines, perhaps the *clowns world* of Robert F. Simmons *could* produce narrative if, for example, the clown's actions could lead to him being in danger of death due to having sex for money. Likewise, we might imagine stories about *Elmer* the taxi driver's destructive scheme to plunge Simon City into chaos under the influence of his powerful lover. Of additional use are Schank's operators for personal relatedness and unexpectedness. While the notion of personal relatedness is an obvious ramification of the case of interactive emergent narrative—indeed, I discussed Schank's thinking in relation to interactivity in Section 3.1.2—unexpectedness may be utilized too. As an exercise, let us take the near story of Simmons and Slocum quoted above and minimally reformulate it according to Schank's ideas:

> Mary was wrestling with a bomb at the liquor bar before John helped her with it. John saw Mary wresting with a bomb at the liquor bar. John went over to help Mary with a bomb before he drew a wire. John and Mary together perished in the blast.

 This modified example, I contend, is undoubtedly a story—I have rendered it *tellable*. Note that I only made three small changes: to serve the absolute interest of danger, ‘bottle' and ‘bottle cork' became ‘bomb' and ‘wire' respectively, and ‘drank the champagne' was changed to ‘perished in the blast' to target the unexpectedness operator taking the absolute interest of death as its argument. [Note: Jonathan Lessard, who is on the reading committee for this dissertation, was not impressed by this modified tale: “Not convinced your updated story is better. It's just more conventional. I adhere more to your earlier argument on ‘art brut' and the weirdness of the naive stories of systems” (personal communication, July 27, 2018). While here I am attempting to make a certain point, one that I will reveal momentarily, I think I probably agree with Jonathan. Sometimes I do not know how to feel.] Now the latter move is one of curation—it pertains to the telling of the story—but unexpectedness at that level must be undergirded by unexpectedness at the level of simulation. If a simulation of Mary and John's attempted detonation at the bar does not include the real prospect of them failing and dying as a result, then unexpectedness at the level of surface expression is probably already a lost cause.

Connecting back to Behrooz's common themes and *Brutus*'s organizing principle, it is possible to target tried-and-true literary themes like underdog stories, narratives of betrayal, and so forth. Again, this is a matter of sculpting the simulation's narrative possibility space—particularly as it is yielded by character actions—to support emergent phenomena that matches these themes. As I will express in Part II of this dissertation, this seemingly obvious notion mostly escaped me for the first several years of my simulation practice. While early on I did model (either explicitly or implicitly) some basic Schankian absolute interests (sex, danger, death), it is at a higher level of dynamic interaction between the actions that produce such intrigue that literary themes such as betrayal emerge. As I note in my discussion of the *Hennepin* simulation engine in Chapter 11, my first authoring goal with that system was to sculpt the possibility space such that *revenge* stories could emerge (and in particular an evocative set of possibilities with arson as *modus operandi*).

Lastly, a simulation may also be crafted to target Ryan's examples of situational tropes rooted in embedded narratives: unsuccessful action, broken promises, violated interdictions, deception, and so forth. Mads Wilbroe and collaborators use Ryan's ideas to discuss a system of theirs [larsen1999interactive] that could not generate tellable events because no alternative possible worlds could be suggested [wilbroe2013virtual]. Indeed, this seems to require some modeling of the internal worlds of characters—their beliefs, their desires, and to also connect back to Elson, the interplay between action and character wellbeing. In Part II of this dissertation, I will discuss my own methodology for modeling these phenomena inasmuch as it has evolved across the development of each of my three simulation engines.

As noted above, several researchers have suggested that unique event sequences are often tellable. Thus, if a simulation yields a large possibility space in terms of emergent event sequences, it might be that the less frequently occurring sequences will be tellable with regard to the simulation domain. Of course, this may not be the case, since the uniqueness of a sequence is in the eye of the beholder: unusual events in the real world are only tellable if the recipient also assigns a low probability of occurrence to those events. [Note: Such probability mismatches are likely at the root of tellability errors in child storytelling: due to inexperience, the child has not yet developed a robust probability model for events, and so events that seem unusual to them (and thus tellable) may seem usual (and thus not tellable) to adult recipients with refined probability models. I think a Bayesian study of tellability could prove fruitful, and indeed such methods have recently emerged in literary analysis [kukkonen2014bayesian].] In the case of a system telling stories about a simulation, this very pattern could occur if the system omnisciently identified an objectively unusual sequence and then told a story recounting it to a human recipient with no refined sense of what is usual or unusual in the simulation's possibility space. While this is a pitfall, if a recipient could in fact develop a probabilistic model for that possibility space, this would be a huge win, since it is trivial to identify unusual sequences.

Indeed, Morteza Behrooz and collaborators have proposed this very method and explored it in the context of gameplay logs from a rummy videogame with chat interaction [behrooz2015remember]. Now Behrooz is exploring how domain features may be used to build semantically richer models of domain interestingness that may be used to identify tellable event sequences [behrooz2015remember]. [Note: As I explain later, in curationist terms, Behrooz's work that I have just referenced concerns the development of *sifting heuristics*. These are heuristics that a *story sifter* module may use to automatically identify interesting event sequences that emerged in the course of a simulation. When such a sequence is identified, the recorded material associated with it is handed off to a *narrativizer* module that constructs a narrative artifact recounting the sequence.] In the case of crafting a simulation, an author can consciously determine what would be interesting in that domain to both support the emergence of interesting material and then omnisciently detect such material as it does emerge. This kind of omniscience (and omnipotence) is a core appeal of building one's own task domain, but typically that comes with the drawbacks of the *toy domain*—these drawbacks are what lead to *interactionist AI* [leahu2008interactionist]. In the case of story generation, however, the drawbacks of toy models are avoided: as Rodney Brooks famously said, “the world is its own best model” [brooks1990elephants, p. 5], and in story generation we might say that *the model is its own best world*. [Note: I owe this rephrasing to Ian Horswill—a former Rodney Brooks student [horswill1993specialization], and a member of the reading committee for this thesis—who used it to characterize my argument (in Section 3.1.1) that simulated worlds are in a sense real.] That is, it is not a proxy for the real task domain, or at least that is not a common construal. [Note: Sheldon Klein realized this in his pioneering early work on story generation. In his first note on “automatic novel writing”, he mentions the appeal of working with a “universe in which the non-linguistic data are completely known and controlled” [klein1969automatic, p. 417]. We might consider, however, what it would mean for a story generator to have our real world as its storyworld—the result would be something akin to *procedural nonfiction*: procedurally generated stories about real-world phenomena. This relates to the concept of *narrativization*, the generation of narrative accounts of raw data, which I discuss periodically throughout this document. Examples of procedural nonfiction include *Terminal Time* [mateas1999terminal,mateas2000generation,domike2003recombinant] and *StatsMonkey* [allen2010statsmonkey].]

#### Toward Personalized Narrative

Finally, though in this case I do not practice what I preach, I think it is worthwhile to discuss the prospect of crafting a simulation that may support Behrooz's notions of topic interest, reminiscence, and implicit familiarity. These dimensions critically rely on subjective characteristics of the listener: who she is, what she likes, what her life has been like, what signs hold symbolic weight for her, and so forth. A simulation whose generated stories could engage a person in such highly personal ways would be an example of *personalized narrative*. In personalized narrative, story generation proceeds from a model of an individual for whom a narrative artifact is specially generated such that the experience of engaging with it may be a reflective one that allows her to mediate on herself and on the world [mateas2016personalized]. My colleague Melanie Dickinson is exploring this area as her primary research agenda here at the Expressive Intelligence Studio. For her, the personalized mode of story generation has the potential not only to produce highly compelling experiences, but to also serve as a facilitator for positive change at the level of individuals, organizations, societies, and humanity [dickinson2017social].

Personalized narrative is still a very young area, however. One line of work has explored the integration of *player modeling* [smith2011inclusive,yannakakis2013player] into interactive storytelling systems. This started with the development of experience managers that consider the history of player actions [goetz1993notes,weyhrauch1997guiding,szilas2003authoring,mateas2003integrating,magerko2006player], but beginning with David Thue's *PaSSAGE* (2008) [thue2007interactive,thue2008passage], systems have sought to also model player preferences [karpinskyj2014video,ramirez2015automated,thue2015generalized,hernandez2015keeping,kim2017computational]. [Note: Though lesser known, an earlier paper by Piotr Gmytrasiewicz and Christine Lisetti (2000) proposed an architecture that would model the player's emotional state and alter gameplay in response [gmytrasiewicz2000modeling]. An intriguing alternative to this kind of personalization is the “inverse user model” [mateas1998subjective, p. 2] of Michael Mateas's *Subjective Avatars* project of the late 1990s [mateas1997computational,mateas1998subjective,mateas1999not,mateas2001expressive]. Instead of learning a model of the player in an effort to fit the story to her preferences and sensibilities, the system subjectively models the player's avatar to have its own internal life and then “actively manipulates [the] user's experience so as to try and make the user feel the same way as a character” [mateas1998subjective, p. 2]. Earlier, in the first tech report on the Oz Project (the tradition in which Mateas was operating as a graduate student), Sean Smith and Joe Bates asked the reader to “Consider worlds where the user is the homunculus inside some other character's head—the user interacts through this different personality and partially assumes it” [smith1989towards, p. 2].] One particularly relevant example of personalized narrative is BBC R&D's ongoing project *Visual Perceptive Media* [gradinar2015perceptive,evans2016creating,cox2018moving], the outputs of which include *Breaking Out* (2012) [forrester2012creation,churnside2016object], an audio drama whose content varies depending on time and geographic location of the listening session, and the short film *The Break-Up* (2015) [frew2017evaluating], which dynamically adapts its music, video, and color grade to the viewer. At the 2017 International Conference on Interactive Digital Storytelling, keynote speaker Pia Tikka discussed ongoing work in the area of *neurocinematics* [hasson2008neurocinematics,tikka2016neurocinematics,tiffee2017rhetorical] that investigates the prospect of cinema that dynamically adapts according to a viewer's brain activity [tikka2017systemic,nunes2017interactive]. One practitioner and scholar in this nascent area is Richard Ramchurn, whose collaborative work on the subject includes several papers [pike2015two,ramchurn2016brain,chamberlain2018design] and the “brain-controlled movies” *#Scanners* (2015) [pike2015scanners,pike2016scanners,pike2016bci] and *The Moment* (2018) [ramchurn2018scanners]. For general discussion on the prospect of personalized media, see these three recent publications [mateas2016personalized,sailaja2017challenges,hook2018facts].

As a future direction, I think practitioners could work to evoke in a particular individual Behrooz's sensations of topic interest (model or target her favorite topics), reminiscence (connect to her past experiences), and implicit familiarity (develop a sign system by which semiotic entities that are cognitively rich for her are mapped onto modeled entities in the simulation domain). The latter effect appears to be the most challenging one to produce, but we have a head start. In the *quantified self* movement [lupton2016quantified], for example, we find an interest taken by individuals to model themselves in a way that could be conducive to the kind of personalized narrative that is envisioned by Dickinson and suggested by Behrooz's notion of implicit familiarity. Furthermore, in the work of thinker and practitioner D. Fox Harrell, we find the application of *algebraic semiotics* to computational narrative [goguen20057,harrell2007griot,harrell2013phantasmal], which is precisely a method for operationalizing the kind of phenomenon characterized by Behrooz's example of a red rose that evokes a visceral sense of familiarity. I can imagine a project like Dickinson's that uses an analogic representational scheme of the kind used in Harrell's work to map between the kind of data captured in the quantified-self movement and the particular entities modeled in a given simulation domain.

### Granularity Extremes

The *granularity* of a simulation pertains to the level of detail of the modeling of simulation entities and processes. In the case of emergent narrative, critical dimensions of granularity include the modeling of space and time, the representation of characters, the breadth of possible character actions, and the level of detail with which a typical character action is modeled. As I will show in this section, works of emergent narrative may be troubled by simulations whose granularities are either too low or too high. [Note: In an earlier draft of this thesis, I used the term ‘fidelity' in place of ‘granularity', since I have heard the former used more often to refer to this concept. However, Noah Wardrip-Fruin brought up a good point in his feedback on that draft: “I'm not sure the word ‘fidelity' is what you want in [this section]. The word, for me, is strongly associated with closeness to an existing phenomenon. But you are also interested in the extent to which novel phenomena (unique to the generative world) are modeled, right?” (personal communication, July 3, 2018). This issue pertains to the curious usage of ‘simulation' to refer to systems that do not model subject phenomena. As I discussed in an earlier footnote, in his paper “Do Videogames Simulate? Virtuality and Imitation in the Philosophy of Simulation” [karhulahti2015videogames], Veli-Matti Karhulahti has decried this usage—which seems to be germane to videogames—because it violates the established scientific definition of the term, which necessitates that a system imitate some real one. By this classical definition, it is natural to discuss the fidelity of a simulation, but this is not the case with the colloquial sense, which I use in this document. This is why I am now using ‘granularity' instead of ‘fidelity'.]

#### *Life and Low Granularity*
Let us begin with some brief discussion of the more intuitive problem: a simulation with too little detail. In the extreme case, it may not be possible to produce anything that would even be called narrative. An example of this might be something abstract like John Conway's *Game of Life* [conway1970game], though the scholar Richard Walsh has argued that the terminology of ‘guns' and ‘gliders'—famous emergent structures in the world of *Life* hacking [adamatzky2010game]—actually constitutes a kind of narration [walsh2011emergent, p. 75]. Nonetheless, by any interpretation, it is probably unreasonable to say that Conway's *Game of Life* produces *good* stories. Later, in Chapter 7, I will critique my first simulation engine, *World*, for its low modeling granularity. [Note: From a different perspective, cellular automata might be deemed extremely *high*-granularity, because the modeled components are so small and modular. This claim makes sense when considered in the context of *artificial life* [langton1997artificial] projects that seek to generate lifelike structures or behaviors that emerge from the low-level patterns of cellular automata [langton1986studying,chou1997emergence,seager2012emergence]. This is a remarkable challenge, however, and the resulting “digital creatures” are typically insectoid [evans2001larger]. Some practitioners in artificial life do attempt to generate narrative, but by eschewing cellular automata to instead start from a higher-order agent representation. A remarkable undertaking in this area is the *autobiographic agents* project of Kerstin Dautenhahn, Chrystopher Nehaniv, Wan Ching Ho, and others [dautenhahn1996embodied,nehaniv1997whats,nehaniv1998semigroup]. Autobiographic agents are autonomous agents who dynamically construct life histories composed of symbolic “memories” of their past experiences; these autobiographies crucially guide (and may be revised according to) activity carried out by the agents as the simulation proceeds [nehaniv1998embodiment]. In this project, autobiographic memory is broadly considered as an adaptive survival mechanism, an idea that may tested through simulations that explore the extent to which agents with autobiographic facilities survive longer than purely reactive agents in simulations with resource competition [ho2003comparing,ho2005autobiographic,ho2005computational,ho2006study,deutsch2008episodic]. Intriguingly, these researchers have also experimented with agents who may *communicate* their autobiographical memories to other agents [ho2004sharing,ho2008computational]. My original dissertation proposal [ryan2016situated] was likewise about autonomous agents who narrativize their ongoing experiences in a simulated world, and in general that notion corresponds to a particular *subjective* mode of curationist emergent narrative. This connects nicely to a fabulous paper by Jonathan Lessard and Dominic Arsenault, titled “The Subjective Interface” [lessard2016character], that proposes a design pattern whereby the ground truth of a simulated storyworld may only be accessed through the subjective interface of a potentially unreliable character.]

#### *Saga II 's Drinking Machine*
Now, let us consider the less intuitive case: simulations with too much detail. Here, it will be illustrative to consider the example of what it currently to my knowledge the earliest work of emergent narrative: *Saga II*. [Note: In fact, to my knowledge this is more broadly the earliest story generator of any kind [ryan2017grimes]. Moreover, I believe it to be the earliest major project in computational art, since the depth and complexity of its computation (and the number of person hours expended in its development) greatly exceed that of earlier projects like Christopher Strachey's love-letter generator [strachey1954thinking,wardrip2011digital] and Theo Lutz's poetry generator [lutz1959stochastische,funkhouser2017ibm]. I first learned of the system in a paper by Mark Sample [sample2013account]. With Gillian Smith, I am currently undertaking an excavation of *Saga II* that has already turned up an abundance of archived materials. (We hope to report on our dig soon.) Finally, Liza Daly has recently produced a playful reimplementation of *Saga II*, which she named *Saga III* [daly2018saga].] In 1960, CBS produced a television miniseries, called *Tomorrow*, that was intended to honor the upcoming centennial of MIT. One episode in this series, titled “The Thinking Machine”, focused on the cutting edge of work being done with computers. As the concept for the episode was being developed, a CBS producer named Thomas Wolfe suggested that the TX-0 computer at MIT could be used to generate a script for a TV western [wolf1960letter,tx0vid]. In response to this challenge, MIT researchers Douglas T. Ross and Harrison “Dit” Morse took on the challenge and spent several months developing a system called *Saga II* [morse1960saga]. [Note: One of the open questions in our excavation project is whether there was an antecedent system to this one—that is, why the ‘II'? Doug Ross, who coined the term *computer-aided design* [ross1960computer] and led development of the Automatic Programmed Tools language [ross1978origins], passed away in 2007. I am currently attempting to make contact with Dit Morse, as I did in a recent project in which I contacted Joe Grimes as part of an excavation of his forgotten 1960s story generator [ryan2017grimes]. A letter to his last recorded address was returned.] Rather than entire television episodes, the system generates choreography for a shootout pantomime between a sheriff and a robber at a hideout. In the actual CBS television show, which was viewed in an estimated 7.7M homes [ross1960note], three of the system's outputs were showcased as professionally produced playlets. (I explain more about this aspect of the project in my discussion of curation in Chapter 5.)

It is interesting to consider the design decisions made during the development of *Saga II*, because they represent what appears to be a first foray into emergent narrative. Ross and Morse had no earlier works to consult in deciding how their project would work, so they were forced to invent an approach to character simulation out of whole cloth. [Note: Even if such an antecedent exists—and please let me know if one does—it is clear that the *Saga II* creators would not have been aware of it. Generally, I think it is more fruitful to consider cases of independent reinvention as though they were cases of first invention—that way, each may provide a different view of the broader phenomenon of that invention. This is the approach I am taking in a history of interactive branching narrative that is currently in development.] In building such a simulation, a primary consideration is granularity, and Ross and Morse chose to represent the modeled action in extreme detail. Specifically, *Saga II* is very detailed in its simulation of characters handling objects: the system keeps track of what is in each of the character's hands, and picking up an object may entail a complex procedure by which objects that are currently held by a character are transferred between a character's hands and or the modeled environment. For example, there is a bottle of whisky in the hideout, which the characters can pick up, move, pour from, and drink from. These affordances drive one of the more interesting features of the system, which its creators call the “inebriation factor”:

> Every time the sheriff or the robber takes a drink—and we programmed opportunities for doing this—the inebriation factor is increased, making intelligent actions a little less probable and unintelligent actions a little more probable. After a few drinks, a man takes longer to do things. After the computer's robber has had several drinks, instead of aiming only once to fire, he often aims three times before firing. [quinn1960press, p. 3]

[Figure omitted]

 As Figure 4.2 illustrates, the procedure for having a drink can be quite complex, entailing a number of steps along the way. [Note: The numbered circles in this diagram represent probabilistic switches that route the simulation processes: when such a switch is encountered, a probabilistic distribution over its branches is computed according to the state of the storyworld, and then a pseudorandom number is generated to determine which branch to probabilistically take. Intriguingly, an interactive mode allows a user to adjust, at runtime, the procedures that determine the various branching probabilities. This is the subject of the creators' only report on the *Saga II* project, an internal technical memorandum written by Morse [morse1960saga]. This mode was developed not for use by interactors, but as a way for the authors to quickly tune system parameters in the midst of impending CBS-TV production deadlines. Thus, one might view this interactive mode as the first authoring tool in procedural narrative. Currently, Yotam Shibolet and collaborators are carrying out a project to ascertain the history of authoring tools for interactive storytelling [shibolet2018framework]; I look forward to reading about their findings.] For instance, if a character is holding a gun in his right hand in the corner of the hideout, to take a drink out of a glass he will have to follow this convoluted procedure: move to the location of the bottle and the glass, transfer the gun to the left hand, pick up the glass with right hand, transfer the gun from left hand to a surface at current location, transfer the glass to left hand, pick up the bottle with right hand, pour whisky from the bottle in right hand into the glass in left hand, transfer the bottle from right hand to the surface at current location, transfer the glass from left hand to right hand, at which point a drink may finally be taken from the glass. [Note: Note that a character may also drink straight from the bottle, which will cause his inebriation factor to increase by a double increment. Intriguingly, the more a character drinks, the more he wants to continue drinking, and so in some generated stories the robber falls into a drunken stupor before the sheriff even arrives at the hideout. In one experiment, the creators turned up the urge to drink, and instead of having a gunfight, the sheriff and the robber spent the scene vying for the bottle. Another humorous output ended with the robber killing the sheriff and having a victory drink, and then eight more, before leaving the hideout without his gun [quigg1960machine].]

While *Saga II*'s detailed simulation of prop handling can lead to some mildly interesting emergent phenomena, the considerable authoring effort entailed was probably not merited. Moreover, this extreme level of detail would likely be unsustainable for a larger-scale storyworld. As Figure 4.2 illustrates, the simple case of a character taking a drink required the authoring of a fairly complex code routine; the other phenomena modeled by the system are driven by procedures authored in this same way. That being said, one could argue that the system is specifically a machine for generating gunfights, and so the level of detail actually nicely fits the domain. That is, gunfights are all about the speed of the draw, and this depends on the ability of a character to quickly handle a gun, and such behavior depends primarily on the handling of objects. Indeed, this rich modeling does lead to some interesting emergent phenomena in *Saga II*: for example, a character could be shot because he had placed his gun in his holder to take a drink. While this is perhaps the case, my sense is that the authoring effort that was exerted to model the details of object handling could have instead yielded emergent phenomena with more narrative intrigue. For example, the inebriation factor itself required little authoring effort—drinks increment a number that alters the probability of taking certain subsequent actions—and produces far more interesting emergent situations than the object handling. While this depends on drinking, which in the implementation depends on object handling, Ross and Morse could have instead modeled drinking at a higher level of abstraction (e.g., a character must just be near the bottle to drink). If Ross and Morse had instead worked primarily at that level of abstraction, they may have yielded an even richer possibility space. That being said, *Saga II* is a remarkable first foray into emergent narrative, and in my estimation an extremely successful project. Later, I will discuss how impressed I am by its level of curation.

#### High Granularity in *Tale-Spin , , and *
It is not hard to name projects in emergent narrative that have attempted to wrangle a simulation whose granularity is probably too high. In a sense, the hallmark of emergent narrative is modeling a storyworld with remarkable detail. Other examples include *Tale-Spin*, whose character actions correspond to the primitives of Roger Schank's *conceptual dependency theory* [schank1975primitive], which means its simulated events may for instance pertain to ‘propel‘ and ‘ingest‘ transformations on objects. At this level of granularity, it takes a lot of authoring to ensure that interesting stories may actually emerge at a higher order than the physical primitives.

In a related example, *Prom Week* [mccoy2011prom,mccoy2013prom] cocreator Mike Treanor has come to believe that the game's underlying social simulation was probably too complex. I asked him about this over email, and Treanor noted that a simpler social schema [mccoy2014social] would have dramatically reduced the project's considerable authorial burden:

> So we had about 5000 “social considerations”. [...] I think the biggest reason we ended up with so many rules is that we weren't as careful about what we put in our social schema [...] We broke things down into 3 binary reciprocal relationships, 3 undirected weighted relationships (networks), maybe like 15 statuses, like 25 traits, 10 or so history labels, and 10 or so cultural knowledge and labels about possession of cultural knowledge. Each one of those features needed to be considered with regard to all combinations of each of the other features [...] So, how could we have gotten away with fewer rules? The easy answer is that I think we could have had fewer traits, fewer statuses, etc. Because of exponential relationships these things have to one another, each one you remove reduces authoring considerably. [Note: Personal communication, April 23, 2018.]

 Of course, a reduction in simulation complexity would likely come at the cost of emergent intrigue, but Treanor believes that the extreme detail of the game's social simulation was overkill with regard to the amount of complexity that players would actually come to comprehend through gameplay:

> But also, once the game was finally finished (with the current interface with the thought bubbles), we actually ended up creating intermediate abstractions to help reduce the amount of things that the player would need to think about. A goal, of mine at least, in Prom Week was for the player to actually think about and care about what [the underlying AI system] *CiF* was doing, and in the end, we needed to reduce it (the contents of the thought bubbles are a synthesis of [social] rules). So, through playtesting, we could have probably figured out what the level of abstraction that the players would be able to actually comprehend was, and had that be what the [social] rules were about (rather than our initial, overly robust decomposition of social reality). [Note: Personal communication, April 23, 2018.]

Of course, there are exceptions to every rule, and here *Dwarf Fortress* is a strong one, since its success has largely been predicated on the extreme granularity of its simulation (a hallmark that is often cited in its various glowing reviews [weiner2011dwarf,winslow2013simcity,hall2014dwarf]). However, its creators' solution to immense authorial burden has been to carry out a thirty-year project [rose2013dwarf], and cocreator Tarn Adams has himself warned of the danger of excessive detail in terms of its exacerbation of authorial burden (and confusion):

> There's no reason to have 50 variables modeling one aspect of a game's behavior if they
don't all have a meaningful impact on the game. Operate at the level of what the player sees
or one layer below. Don't get carried away and create a lot of worthless noise. Not only do
uselessly complicated systems hinder tuning, but they can have a paralyzing effect during
development. Use complexity only where it's needed, otherwise strive for simplicity. [adams2015simulation, p. 520]

In my own work, I have struggled at various points with both granularity extremes, as my self-critiques in Part II will show. Granularity poses a tough design challenge, since homing in on the right level is a process of discovery that entails a lot of iteration and tuning. In his email, Mike Treanor remarked as much following his explanation of *Prom Week*'s level of detail:

> But there was no way we could have known any of that at the time! [...] I don't know, we just gotta make a lot more of these things in order to make any sense of authoring for it. [Note: Personal communication, April 23, 2018.]

### Low Modularity

Some character simulations are not modular enough. Though this pitfall may also manifest in adverse surface effects, as I will explain, it relates primarily to the authorial considerations of simulation crafting. Here, I am specifically concerned with the modularity of the procedural content that defines how the storyworld is represented (for instance, how characters are modeled) and how it may evolve (namely, character actions). Modularity in this sense increases as monolithic chunks of procedural content are broken into smaller pieces that may be recombined in a variety of ways.

#### Klein's Misunderstood System

Modularity is low, then, when the simulation's procedural content is structured as a collection of relatively monolithic chunks. As an example of this, we might consider the murder-mystery generator developed by Sheldon Klein and his students in the early 1970s, the modularity issues of which have been cited as the impetus for the tidal shift in story generation that lead to the predominance of planning techniques. [Note: Klein's system has for decades been known as the earliest effort in story generation, but in a recent paper [ryan2017grimes] I reported on an excavation of three forgotten earlier efforts: the aforementioned *Saga II* of 1960 [morse1960saga], Joseph E. Grimes's unnamed folktale generator of circa 1963 [grimes1963ibm,grimes1965linguistic,grimes1965computadora,grimes2017translation], and Robert I. Binnick's unnamed folktale generator of 1969 [binnick1969application, p. 27].] As Noah Wardrip-Fruin explains, “the field has generally followed Meehan's critique of the Klein's system: it was made up of explicit chunks of action, with a path through these chunks selected randomly” [wardrip2009expressive, p. 116]. [Note: Note that there are actually two components to this critique. The first relates to the modularity of the procedural content, referred to by ‘chunks of action,' and the second to a lack of abstraction in designing that content, denoted by ‘explicit'. In this section, I discuss modularity, and in the next I discuss the related issue of abstraction.] Indeed, it is not hard to find recapitulations of this criticism by contemporary scholars [liu2002makebelieve,jaya2010novel]. Though there is some veracity to it, the origin of the critique seems to be rooted partially in misunderstanding.

While many have characterized Klein's system as a story grammar or template-based system [yazdani1982write,yazdani1989computational,steiner1992graphic,faas2002virtual], that is very far from the case, as I will now show. [Note: In a *story grammar*, stories are generated by terminally rewriting the start symbol of a context-free grammar whose nodes correspond to either plot events or segments of a natural-language narrative telling [rumelhart1975notes]. As you will read momentarily, this is a ridiculous characterization of Klein's system. I have a few theories for how it may have emerged, though. First, Klein illustrates the workings of his system in figures that use *Backus–Naur form* [klein1971program, pp. 6–8], which is a grammar-like notation that is commonly used to define the operation of programming languages [mccracken2003backus]. To the uninitiated, this could be seen as an indication that Klein's system has the power of a context-free grammar, but it is in fact far more powerful. Second, Klein was a participant in the *text grammar* movement of the 1970s [klein1975meta], which sought to develop structural accounts of meaning at the level of discourse (thus, ‘text' as a discursive object, not a string of symbols), as an opposition to the Chomskian fetishization of the sentence [van1972some]. This movement had its locus in Europe, and I believe unfamiliar American scholars may have assumed it was concerned with story grammars. Lastly, story grammars were the subject of harsh critique by scholars whose associates sought to distance themselves from Klein's work: story grammar's co-executioner Robert Wilensky [black1979evaluation] was a labmate of Natalie Dehn and Michael Lebowitz [dehn1981story,lebowitz1984creating], who led the field of story generation in the early 1980s and who, in seeking to substantiate their own contributions, may have found it convenient to reductively conceptualize Klein's system as a kind of story grammar. Lastly, I will note that while the misconception is predominant, not all scholars have succumb to it [ryan1987heuristics].] More aptly, the architecture should be described as a *production system*. [Note: Roger Midmore has argued that Klein actually invented the *blackboard architecture* [midmore2014interpretation, p. 9], which is related to the notion of a production system, though Klein's inception was in an ancestor to this system [klein1964dynamic,klein1965some]. Production systems are believed to have been first described by the mathematician Emil Post in 1943 [post1943formal], building off his earlier work [post1936finite], though Paini's ancient generative grammar (with its rules about rules and other mechanism) is a notable antecedent [staal1965context,katre1989astadhyayi,bhate1991panini]. In any event, computerized production systems were just emerging [davis1975overview] as Klein began to incorporate them into his own technical practice [klein1964dynamic,klein1965some,klein1966historical]. He does not use the standard terminology of production systems, which suggests that Klein may have independently invented the concept. For the sake of clear explanation, I will use the conventional terminology here.] In a production system, the fundamental unit of procedural content is the *production rule* (or just *rule*), which is a package comprising two components: a list of *preconditions* specifying aspects of the system state that must hold in order for a rule to be executed, and a list of *effects* (also called *postconditions*) that define how the system state will be altered if the rule is ultimately executed. In a production system, the portion of system state that may be altered by production rules is represented as a knowledgebase containing *facts* about the world, which is called the *working memory*. Typically, a rule is tested by evaluating each of its preconditions—if all of them hold, the rule is *triggered*, and subsequently all of its effects will be *executed* to update the working memory. Klein, however, introduces an additional wrinkle. In his system, authors associate preconditions with *weights* (probability increments or decrements), and the result of testing a rule is a probability of the rule actually being triggered. That is, as each precondition is evaluated, the sum probability of triggering the rule is either incremented or decremented according to the weights associated with the preconditions. [Note: Note that preconditions in Klein's system may bundle multiple predicates, with variable binding across the predicates. For example, the rule ‘GEORGE KNOWS & $$ HATES & $$ DISLIKES Y‘ [klein1971program, p. 14]—which may be read as ‘George knows and does not hate and does not dislike $Y$'—is a composite of the three predicates $knows(George, Y)$, $ hates(George, Y)$, $ dislikes(George, Y)$, with $Y$ bound consistently to the same entity across the predicates. An $or$ operator may similarly be used. Additionally, each precondition is more precisely associated with two probability increments or decrements: one that is used when the precondition holds and another that is used when it does not. Finally, effect execution can be made deterministic, rather than probabilistic, by simply using numbers outside the range $0.0-1.0$. That is, if the weight $10$ is associated with a precondition, and that precondition holds, then the effects in the rule will certainly be executed (unless, of course the author also associates huge negative weights with other conditions in the package).] Once a sum probability has been computed, a pseudorandom number (between $0.0-1.0$) is generated, and if it is less than the computed probability (treated as a floating-point number), the rule is triggered, and thereby all its effects are executed. By executing effects, new facts may be added to the working memory and existing facts may be removed. [Note: To be a precise, in lieu of removing a fact, a new fact representing its negation is added to the working memory. Additionally, the working memory keeps track of how long it has been since each fact was asserted, which allows for preconditions to check how much story time has passed since an action occurred or a stative was asserted (or negated). Because the deletion of a fact is represented by the addition of a new fact that negates the original one, the system can also reason about the duration for which statives were maintained, including the specific times (diegetic timesteps) of onset and termination. This affords both powerful authorial control (at authoring time) and expressive system reasoning (at runtime).]

By hacking his production system to work probabilistically, Klein's architecture cleverly approximates the expressive power of *utility-based action selection* [maes1989right,mark2009behavioral], while still keeping within the computational limitations of the day. To illustrate this, let us compare it to *Comme il Faut* (or *CiF*, pronounced ‘siff') [mccoy2014social], the utility-based AI architecture that underpins the videogame *Prom Week* [mccoy2011prom,mccoy2013prom]. While in Klein's system precondition evaluation derives a probability of executing a list of effects, in *CiF* precondition evaluation derives a utility of taking a prospective action. (For our purposes, executing a list of effects and taking an action may be treated as being equivalent.) In *CiF*, the utility of all candidate actions is computed before an action is taken (the top-scoring one), which provides guarantees that the most fitting action (as implied by the authored rules) is always taken. This guarantee necessitates the evaluation of potentially thousands of preconditions, which in Klein's day would not have been feasible. In one paper, Klein warns of the danger of evaluating even a few dozen preconditions:

> Then, for each X and Y pair, evaluate the [preconditions] for this rule. [...] Note that if the class specified for the X variable contains 8 objects, and that for the Y variables contains 5 objects, the [preconditions] for this rule will be evaluated 40 times. Such nested loops can be very time consuming and should be used with caution. [klein1971program, pp. 13–14]

Thus, the naive solution, which is moreover a common operating mode for production systems, is to trigger the first rule whose preconditions all hold. But in return for its speed gains, this method sacrifices any hope of the optimal action being taken, since the corresponding production rule may not be the first one whose preconditions are evaluated. Klein, however, designed an intermediate approach that uses author-defined probabilities to ensure that more fitting actions (given the current state of the storyworld), though not guaranteed, will be more *probable*: instead of deterministically triggering a rule whose preconditions hold, Klein's execution probability allows his system to potentially bypass the rule to choose another one whose preconditions also hold. This provides an approximation of utility without necessitating the evaluation of all rules before deciding which one should be triggered. As explained in Section 11.2.7, I utilize this same technique in my current simulation engine, *Hennepin*, since its expansive number of characters makes it susceptible to the computational limitations of our current day. [Note: The earliest writing on utility-based action selection in story generation that I am aware of appeared on the ‘rec.arts.int-fiction‘ newsgroup [fisher1992use] in 1991, where it was proposed by Robert Taylor Fisher, an undergraduate at the California Institute of Technology who was working on interactive storytelling at the time [fisher1991my]. An early implemented example by Brad Rhodes and Pattie Maes, based on Maes's *behavior network* architecture [maes1989right], appeared later in 1995 [rhodes1995stage].]

Additionally, in Klein's architecture rule effects may alter the operation of the production system itself. Rules are associated with *frequencies*, where rules with higher frequencies are tested more often as story time progresses, and there is also a *priority* policy that determines the order in which rules will be tested. Intriguingly, the effects of executed rules may *disable* (or re-enable) other rules—disabled ones will not be tested—or modify their frequencies and priorities [klein1973automatic, p. 15]. [Note: Additionally, an author may organize rules into groups, which allows for the modification of activation statuses, frequencies, and priorities at the level of groups. Rule effects may also dynamically reconstruct the organization of rules into groups by forming new groups, removing groups, and changing the group memberships of individual rules. Again, this affords immense authorial power, in this case to a level that surpasses that of nearly all systems, historical or contemporary, that I am aware of.] This dynamic reformulation of the operation of the production system according to the effects of its own rules is a kind of *computational reflection*, which is a design pattern that characterizes computer programs that can read and modify their own code [smith1982procedural]. Later examples of the use of this technique in computational narrative include extensions to the Oz Project's Hap language [bates1992building,sengers1998thing], *Facade*'s ABL language (whose *meta-behaviors* use reflection similarly to how it is used in Klein's architecture [mateas2002interactive, pp. 93–96][mateas2002behavior, p. 41]), and the Gertie language developed at Zoesis in the 2000s [loyall2004system, p. 62]. [Note: As Michael Mateas notes in his dissertation, “powerful features such as reflection, which edge towards ‘you can do arbitrary things', are paradoxically useless without idioms providing guidance on how to use the feature” [mateas2002interactive, p. 94]. One such idiom in Klein's framework is the deactivation of rule groups that are associated with beats in the story that have already passed. For example, after the murder occurs, the rules for playing tennis or going to a pub may be deactivated. This both saves on computation time, since those rules will not be needlessly evaluated after becoming obsolete, but it also provides an *authorial affordance* [mateas2002interactive, pp. 125–126] that enables an author to conceptualize higher-order patterns that group character behaviors according to notions that might be called scenes or *beats* [mckee1997substance] or *social practices* [evans2008re,evans2014versu,treanor2016framework,evans2016computer]. In later work, Klein further explored reflection by introducing effects that not only reformulate the operation of the production system, but *generate* entirely new rules. In turn, these generated rules could contain effects that generate more new rules, and so on. Klein called this concept *meta-compiling* [klein1975meta]. It is worth noting that Klein's system may not have been the first in computational narrative to use reflection: due to its vintage, *Saga II* likely used the method merely due to being programmed in assembly for a computer with limited memory. In later examples, however, reflection obtains also at the level of the *rhetorical machine* of an AI system [mateas2003semiotic], where it becomes a human-understandable computational move that may underpin specific authorial affordances, such as Klein's rule-group control.  Design patterns like reflection feature heavily in the aesthetic space of *weird languages* [mateas2005box].]

Clearly, Klein's system is no story grammar. So, how does it generate stories? In this architecture, the working memory is the very representation of the storyworld and its accumulated history. [Note: Later story generators that are also driven by production systems include *Telltale* [simmons1979rule,correira1980computing] (by Alfred Correira and Robert F. Simmons, who was Klein's 1960s collaborator [simmons1962toward,klein1963syntactic,klein1963computational,simmons1964indexing]), which also seems to be misunderstood as a story grammar [smith2001representation, p. 303][wilcock2005automated, pp. 8, 18][sennott2012quicksilver, p. 3], and an obscure 1986 system by Marie-Laure Ryan [ryan1986structure]. Additionally, the aforementioned *Prom Week* [mccoy2014social] utilizes production systems [mateas2013architecture], as does *The Sims 3* [evans2010modeling] and *Versu* [evans2014versu], for each of which Richard Evans served as an AI lead.] In a beautiful expression of how nascent the concepts of story generation and production systems were at the time, Klein calls such storyworlds “stochastically modifiable semantic models of arbitrary universes”, and the stories that are generated thereby “reports on the status of the modelled universe” [klein1971program, p. 1]. To generate such a report, the system uses special rules that match against collections of facts in the knowledgebase to target other rules in a context-sensitive grammar, whose execution produces strings that render the storyworld facts in natural language. Because facts in the knowledgebase are represented as triples of the form ‘(subject, relation, object)‘ and tuples of the form ‘(subject, attribute)‘, the working memory yields a *semantic network* that represents entities (e.g., characters, locations, props) as *nodes* and associates them with their attributes and also connects them to one another according to relations (e.g., $loves$, $hates$) that are represented by *edges* in the graph. [Note: It appears that Klein and Robert F. Simmons, while working together at System Development Corporation (the first software company [baum1981system]) in the early 1960s, independently invented the notion of a semantic network [klein1963syntactic], though they were preceded in this by the earlier projects of Richard H. Richens [richens1956preprogramming] and Margaret Masterman [masterman1961semantic,sowa1992semantic].] Additionally, authors may specify composite nodes in the network that group subgraphs that correspond to higher-order concepts. For example, Klein notes how the concept of a series of events causing another series events could be reasoned about to construct a simple triple expressing $implies(A, B)$, where A is a composite node containing the subgraph corresponding to the first series of events and B is a subgraph constructed accordingly for the second series [klein1973automatic, pp. 5–7]. I will revisit this aspect of the architecture in Section 4.2.3 when discussing the notion of selecting a subset of simulated material for surface expression.

[Figure omitted]

#### Monolithic Murder Mysteries

Klein's architecture probably appears to be quite modular at this point. But while it certainly *supports* the authoring of very modular procedural content, the content that was actually authored for the murder-mystery generator lacks modularity, and in two particular ways. To explore this, let us consider the diagram for the system, shown in Figure 4.3, which was supplied by Marie-Laure Ryan in a fabulous 1987 paper [ryan1987heuristics, p. 514]. For something to be modular, it must have two characteristics: it is made up of small pieces, and the pieces can be recombined to construct more complex higher-order structures. As I will explain momentarily, Klein's system illustrates how the procedural content authored for a simulation may lack modularity in either or both of these senses.

Ryan's diagram shows that the procedural content that governs character behavior is made up of pieces, but some are more monolithic than others. [Note: As I also note in the figure caption, this diagram is not consistent in the fidelity with which it represents portions of the narrative possibility space. Specifically, while Ryan has taken care to represent all the major branch points preceding the murder, her depiction of the murder beat itself collapses a series of possibilities into one composite node. In fact, there are six ways that the murder can play out [klein1973automatic, p. 67], as determined by the particular motive (of which there are also six) that the murderer has taken on, which is itself determined by how the earlier simulation phases turned out. In general, Ryan's diagram does not fully capture the degree of coordination between emergent actions and later ones. Later, in Section 4.1.6, I discuss this notion of simulation feedback as perhaps the most critical design concern in emergent narrative.] For example, while the modeling of character flirtation and a subsequent tryst spans multiple high-level events that may themselves recombine to yield a fairly rich possibility space, the questioning by the cops is a single monolithic node that feeds back onto itself in an act of repetition. A more modular approach would also break down the police investigation into its constituent parts, ideally in such a way that the parts could be recombined to yield meaningful narrative variation. One possibility, however, is that the program is meant to model the possibility space of an existing fiction genre, the murder mystery, and the tropes of that genre imply that it is not important to richly model (in this case) the nuanced possibilities of a bungling police investigation. This is a fair argument—and the particular monolithism under consideration here may also have been due to a lack of authoring resources, as Section 4.1.5 notes—but I believe that the authors of simulation engines for emergent narrative should strive toward modularity, since that yields greater emergence, which amplifies the aesthetic qualities of the form as outlined in the previous chapter.

The other component of modularity, besides breaking monolithic chunks into smaller pieces, is the capacity for those pieces to be recombined into a variety of higher-order structures. In Klein's system, we find cases where potential narrative beats do not branch out to enable one or more possible ramifications, but instead feed forward (more or less deterministically) into another narrative beat. For example, in the diagram shown in Figure 4.3, this feedforward pattern occurs in the linking of the nodes marked ‘x and y get out of bed at night and meet in hall‘ and ‘x and y go to secret room‘. When a sequence of chunks is connected in this way, like beads on a string, the monolithic structure that was broken into these chunks is reconstituted—it has simply become a monolithic sequence, rather than a chunk. A more extreme example of this is the node called ‘cops question w‘, which feeds back onto itself to produce the monolithic sequence of the same event recurring over and over again.

#### A Combinatorial Magic

Again, modularity is critical when it comes to the procedural content of emergent narrative, because it is what enables the rich possibility spaces that undergird the aesthetics of the form. A shining example of modularity in emergent narrative is the game *Caves of Qud*, by Jason Grinblat and Brian Bucklew [grinblat2010caves,grinblat2017subverting]. In a recent book chapter, Jason Grinblat has articulated the appeal of modularity:

> Modularity is the use of discrete units, called modules, to assemble larger structures, which we'll call gestalts. Our motivation for investigating modular design comes from the combinatorial magic that modularity conjures. We design a few modules, along with an assembly mechanism, and we inherit a plethora of gestalts for free. And gestalts are what matter. They're the artifacts our players encounter and care about, be they dynamic puzzles, dungeon levels, or dialog trees. Modularity multiplies our work, often bearing novel results in the process. [grinblat2017designing, p. 29]

 Tarn Adams has likewise endorsed the virtues of modularity, using an example from the development of *Dwarf Fortress*:

> If you just focus on the end result, it's not clear what's required to make it look and work correctly. Instead, break down and understand the system you are modeling in terms of its basic elements and interactions. Not only will you develop a richer interplay of objects, but certain problems solve themselves. As a simple example, when creating terrain, it is tempting to spawn particular biomes or allow a fractal to directly define the biomes. However, Dwarf Fortress achieved much better results by handling fields separately: temperature, rainfall, elevation, drainage, etc. The interplay of those fields determined the final biome, resulting in a more natural, internally consistent solution. [adams2015simulation, p. 520]

In total, the case study of Klein's system provided in this chapter shows that a modular architecture, though probably necessary for strong emergent narrative, is only half the battle: the procedural content authored for that architecture must itself be modular as well.

### Lack of Abstraction

While modularity pertains to the granularity and combinability of procedural content, a related issue concerns the level of abstraction that is used to define such content. Here, let us return to Noah Wardrip-Fruin's critique of Klein's system, in which he compares it to *Tale-Spin* in a particular way:

> But the field has generally followed Meehan's critique of Klein's system: it was made up of explicit chunks of action, with a path through these chunks selected randomly. So, for example, Klein's system includes a ‘rule for people arriving at George's living room'. Meehan's system, on the other hand, contains general rules for character movement (and reasons for movement) that operate no matter which spaces are available—living rooms, bedrooms, caves, or meadows. [wardrip2009expressive, p. 116].

 Here, Wardrip-Fruin is articulating a distinction between the two systems in terms of *abstraction*: he argues that Klein's system utilizes concrete chunks of procedural content that are explicitly associated with certain characters and places, while *Tale-Spin*'s procedural content is defined abstractly so that it may be utilized for any character in potentially any context.

#### Klein's Concrete Elements

As the discussion in Section 4.1.3 illustrates, Klein's procedural content does in fact utilize abstraction, with rules whose predicates may match against arbitrary features of the storyworld (and even aspects of the architecture itself, as a form of computational reflection). To be clear, though, there are also a few rules that reference specific concrete entities. In the example that Wardrip-Fruin cites, a rule determines which characters the host George will invite to his cocktail party [klein1971program, pp. 14–15]. This is not, however, a concrete plot event that is simply triggered by the rule. Rather, whom George will invite depends on aspects of the storyworld as defined in the initial state, namely his affinities for the other characters, and so changing the initial state may cause this rule to be triggered differently, which may totally change the story. [Note: In this sense, Klein's system actually works a lot like *Tale-Spin*, which Meehan described as being interactive due to its user affordances for asserting facts about the storyworld [meehan1977tale].] The reason George is explicitly referenced in the rule is that the setting of the initial murder-mystery domain (described in a preliminary report on the project) is a cocktail party at George's house, and so every generated story begins with George inviting his guests [klein1971program, p. 27]. Likewise, Klein's revised story domain (referenced in all subsequent reports on the project) also has a concrete setting—Lady Buxley's countryside manor [klein1973automatic]—and some concrete rules pertaining to specific characters and locations within the manor.

Thus, while Klein's system features abstraction, it also features concrete elements, and this is what underpins the fundamental critiques of the system (though, as I have shown, these are largely misguided). [Note: *Tale-Spin*'s storyworlds also have concrete characters and settings at runtime, since generally these elements are defined at authoring time, just as in Klein's project. The major difference is that *Tale-Spin*'s interactive mode cleverly interleaves story authoring and story execution by asking the interactor to define aspects of the storyworld (e.g., one character's affinity for another) as they are needed to evaluate conditions on activated planning operators. In 1986, Laxmi Gupta amplified this aspect of *Tale-Spin* in a reimplementation of the system [gupta1986variant]. Later still, Michael Gardner further extended this interaction pattern to build a mixed-initiative authoring tool called *Multistory* [gardner1991expert]. As I write this, Ben Samuel is developing a system called *Writing Buddy* that pushes the idea even farther [samuel2016design]. Another related prospect for procedural narrative is the system itself filling in these deferred gaps automatically, in a way that would best serve the emerging narrative. This approach was explored by another early system by Masoud Yazdani, called *Roald* [yazdani1989computational]. More recently, is has been articulated as *late commitment*, a technique by Ivo Swartjes and his collaborators that takes its inspiration from improvisational acting [swartjes2008let,swartjes2008emergent, swartjes2009late]. As I will argue at some length in Chapter 5, I find these approaches to be highly intriguing, but I believe they represent a departure from emergent narrative and its core aesthetics.]

#### The Aesthetics of Abstraction

Note that Klein and his students could have replaced the concrete elements with more abstract procedural content. In this configuration, the stories would not all take place at the same setting—George's house in the first domain, and Lady Buxley's manor in the second—but instead rules would be authored for generating a setting. Likewise, the characters in the stories would not be the same every time, but instead they would be generated according to a different set of authored rules. [Note: As noted above, *Tale-Spin* also features concrete storyworlds and characters, but some of the authoring of such entities is offloaded to a human interactor. What would it look like if the system itself generated such material? Since *Tale-Spin* is all about planning, such a reformulation would probably use the same approach for this task. But while *Tale-Spin* only used planning for character behavior, this would constitute its utilization for *author* behavior. This is the basis for *Author*, a follow-up system that was proposed by Meehan's labmate Natalie Dehn [dehn1981story,dehn1981memory,dehn1989computer]. Dehn's ambitious project developed alongside her advisor Roger Schank's theory of *dynamic memory* [schank1982dynamic], but *Author* does not appear to have been implemented [dehn1989computer]. (I have also gathered this from a phone call with Roger Schank, who did not recall the system being completed [personal communication, June 28, 2018].) Meanwhile, this intellectual period of the Yale AI Project also produced some of the earliest work on *case-based reasoning*, namely Janet Kolodner's *Cyrus* [schank1980conceptual,kolodner1983reconstructive] and Michael Lebowitz's *IPP* [lebowitz1980generalization,lebowitz1983memory]. Shortly thereafter, Lebowitz developed *Universe*, an early example of how the integration of authorial and character planning might work in a story generator [lebowitz1983creating,lebowitz1984creating,lebowitz1985story,lebowitz1987planning]. Though conventionally unrecognized, this territory was actually first explored by Masoud Yazdani's little-known *Roald* system [yazdani1982write,yazdani1983generating,yazdani1987reviewing,yazdani1989computational,yazdaniundatedroald], as Lebowitz himself acknowledges via citation [lebowitz1985story, p. 485]. Later, Erik Mueller and Scott Turner—both students of Michael Dyer [dyer1985artificial], who was himself a student of Schank's student Wendy Lehnert [dyer1982depth]—produced two other important systems in this tradition, *Daydreamer* [mueller1985daydreaming,mueller1990daydreaming,mueller2018update] and *Minstrel* [turner1985minstrel,turner1986thematic,turner1993minstrel].] In this way, the story generator would become *roguelike*, and thereby the aesthetics of emergent narrative, which I outlined in Section 3.2, would be amplified. Let me explain how.

When setting and characters are concretely modeled, every generated story is about the same people in the same places, and the result may feel more like fiction than nonfiction. Moreover, attendant to this kind of implementation may be a sense that with each newly encountered story an interactor is converging on a fuller understanding of a single monolithic, though multifaceted, story. These sensations are in direct opposition to the pleasures of emergent narrative, and as a result such concreteness works to inhibit its core aesthetics. By authoring characters, the aesthetics of the unauthored are damaged, and in featuring the same characters across everyone's stories, the aesthetics of the personal are undermined. One component of the aesthetics of the vast is the magnitude of the possibility space not just within a given simulated storyworld, but across all the storyworlds that a simulation may produce—because this higher-order possibility space is constrained when concrete characters and settings are used, the aesthetics of the vast are inhibited in turn. Finally, when a reader or interactor finds that the characters of her world persist across simulation instances, the aesthetics of the ephemeral are diminished.

#### Along the Abstraction Continuum

In procedural narrative, the extreme case of low abstraction is probably represented by something like Arthur Blanchard's *Movie-Writer* [blanchard1916movie], which is the earliest project in the area of which I am currently aware. Patented in 1916, *Movie-Writer* is a small device that enables the semiautomatic production of outlines for silent-era movie scenarios. As Figure 4.4 illustrates, this works by a simple process of recombination, where concrete chunks of content (each being a single word) are mechanically juxtaposed. [Note: Related mechanical devices for procedural generation include Henrietta Rosa Montague's *Name Selector* (1947) [rosa1947name] and Norvell E. Von Behren's device for use in baby naming (1968/1971) [von1968word,von1971word]. For a history of procedural generation in analogue games, see Gillian Smith's fabulous paper [smith2015analog]. Moving back to procedural narrative, we might consider the most extreme possible case of concreteness. To my imagination, this would be something like a rolodex of *complete* stories, the ‘procedure' of which would be actuated by a human shuffling through the rolodex (or simply revealing the next card). Here, there is no recombination, or even any generation at all, but instead a pseudorandom *selection* of complete stories. In fact, this rolodex would be procedurally equivalent to a book of short stories that a reader may page through pseudorandomly, or even in order. As such, this case does not even feel like procedural narrative. Indeed, it seems that some amount of abstraction is necessary for the form to even obtain; in the case of *Movie-Writer*, such abstraction comes in the form of its word groupings and the scenario template that houses them in an ordered array of slots. *Movie-Writer* feels generative, but it could also be seen as an example of selection: the wheel positions index the stories, and through turning its wheels, one of its 200 stories is revealed. Why does it feel generative when the stories are broken into chunks? Indeed, all procedural generation may be viewed as being essentially a matter of selection—that is, selection of one artifact from the generator's possibility space (this is the sense evoked by the AI term of art *search* [korf1987planning,togelius2011search]). From this perspective, all a system's generable outputs exist all along in the space, and through some procedure—computational, human-driven, or hybrid—a particular output is *retrieved*. Taken to its extreme, this notion could lead us to view computation itself as a matter of retrieval: a computing machine is a device for retrieving data instances that already exist in the space of possible data that is yielded by the device's electromechanical system. I return to this ontological position in footnote Section 87 of Chapter 4.] A later 1930s project by Wycliffe A. Hill works on the same principal by allowing interactors to generate storylines using a device called the *Plot Robot* (alternatively, the *Plot Genie*)—a paper number wheel that works as a pseudorandom number generator—to recombine character and event descriptions found in a series of books [hill1931action,hill1931romance,hill1931plot,hill1936comedy].

[Figure omitted]

Moving toward the other side of this continuum, we find *The Uranium Shipment and the Space Pirates* (1955), an extremely interesting artifact that Brian Moriarty has identified as the earliest known work of digital interactive storytelling [moriarty2015sing]. Specifically, it an example wiring of the *Geniac* kit [berkeley1956small] that is included in its instruction manual [garfield1955geniacs, pp. 30–31]. Marketed as “an exciting interplanetary chase pitting yourself against the machine” [geniac1956ad, p. 16], this *Geniac* wiring allows a human to select initial conditions for a procedural sci-fi story by turning a set of discs on the face of the kit. By the power of combinational logic, each input configuration automatically triggers a particular ending to the story, which is indicated to the interactor by a lightbulb display. [Note: Moriarty's student Zackery Mason has developed *Virtual Geniac* [mason2016historic], a virtual simulation of the *Geniac* device that allows one to experience of *The Uranium Shipment and the Space Pirates* today.]

Further along the abstraction continuum, we encounter *story grammars* [rumelhart1975notes]—and well-documented antecedents such as Propp's folktale model [propp2010morphology] and William Wallace Cook's *Plotto* [cook1928plotto], as well as several obscure earlier ones [lindsay1916art, pp. 171–189][polti1916thirty,hill1919ten,palmer1922photoplay]—where abstraction appears in the form of nonterminal symbols (representing plot events or other story material) that may be rewritten in different ways depending on the grammar's production rules. Moving further still, we find *roguelike* configurations, where characters and other story elements are procedurally generated for each runtime instance. Examples of such a configuration include *Dwarf Fortress* [dwarf], *RimWorld* [rimworld], *Caves of Qud* [grinblat2010caves,grinblat2017subverting], and my own simulation engines *World*, *Talk of the Town*, and *Hennepin*, which I describe at length in Part II of this dissertation.

To produce something at the far end of this continuum—an antithesis to the approach of *Movie-Writer*—would probably require something like *program synthesis*, whereby the procedural content that defines the rules of the simulation would itself be generated [manna1971toward]. [Note: Perhaps this is what Sheldon Klein was working toward with his increasing exploration into computational reflection. While the most famous work on program synthesis occurred in the mid-1980s [manna1986deductive], it has a much longer history. Indeed, the task is sometimes called *Church's problem* [thomas2009facets], due to Alonzo Church's 1957 proposal for automatic circuit synthesis [church1960application]. Even earlier, in 1954, the term ‘automatic programming' was applied to early Fortran compilers, since at that compilation was seen as a kind of program synthesis [balzer198515, p. 1257]. When Harrison “Dit” Morse undertook development work on *Saga II* in 1960, he developed a series of *macrocode* layers that would each compile code at one level into lower-level code at the next [tx0vid, 3m0s]. At the time, this work was at the cutting edge of software engineering practice.] Steps toward this include the synthetic behavior trees of *Facade* [mateas2004behavior] and the modular entity system of *Caves of Qud* [grinblat2017designing].

#### Retargeting in *Comme il Faut and Klein's System*
I would like to end this section by discussing a particular method of abstraction in procedural narrative, and in software development more broadly, called *retargeting*. In character animation, this term refers to a technique whereby abstracted animation patterns may be authored independently of any given character rig and then dynamically retargeted to a specific rig at runtime [gleicher1998retargetting].  While the notion originated in computer graphics, it may be more generally conceived as a software design pattern [gamma1994design], and there is a clear application to procedural narrative: procedural content, such as character actions, that may be retargeted to various characters or other storyworld entities. Here, it appears that the term was first applied to this domain by the creators of *Prom Week* [mccoy2011prom,mccoy2013prom], or more precisely, its underlying *Comme il Faut* (*CiF*) architecture [mccoy2014social]. As Michael Mateas and Josh McCoy explain in a book chapter dedicated to the topic, this is how retargeting works in *CiF*:

> Multicharacter social exchanges are explicitly represented separately from any specific characters. Given a cast of characters, with traits and social state declaratively represented, social exchanges are retargeted for specific characters. [mateas2013architecture, p. 517]

 In *Prom Week*, this means that both character *social exchanges* (actions that cast multiple characters in specific roles [mccoy2014social, p. 100]) and their *instantiations* (concrete performances of those exchanges that couple natural-language dialogue and character animations [mccoy2014social, pp. 100–101]) may be retargeted to any character. To enable such retargeting, an author defines (using condition logic) the social and narrative patterns in which the procedural content may be used, as well as the rules for rendering concrete instantiations of that content for a given set of characters (or, more broadly, for a given emergent context). Note that Klein's system also features this kind of retargeting, with actions that may be taken by variant characters pending dynamic aspects of the storyworld. Interestingly, Klein and his students intended for the “preselected hero” Dr. Hume to always be the character who solves the mystery [klein1973automatic, p. 70], but this is not possible in all storyworlds: depending on his personality and affinities for other characters, Dr. Hume may end up becoming either the murderer or the victim. When this happens, the procedural content by which the crime is solved will be retargeted to another randomly selected character [klein1973automatic, p. 70].

#### Roguelike Retargeting

There is a more extreme case of retargeting than that which is utilized in *Prom Week* and Klein's system: instead of retargeting behavior to any preauthored character, behavior may be targeted to any *possible* character. This is the roguelike approach to retargeting, and it is at play in projects such as *The Sims* [maxis2000sims], *Dwarf Fortress* [dwarf], *Caves of Qud* [grinblat2010caves,grinblat2017designing], and my simulation engines that are discussed in Part II of this dissertation. As I explain in Chapter 11, in authoring procedural content for my most recent system, *Hennepin*, I conceptualize narrative systems like the one developed for Klein's project as *story machines* that may be retargeted to arbitrary characters at arbitrary times and places. Indeed, retargeting may be conceptualized at different levels of abstraction, and at different levels of hierarchy, with regard to a system's procedural content.

#### Authorial Leverage and Authorial Affordances

Generally, retargeting—and the broader notion of abstraction—is useful in emergent narrative because it provides both *authorial leverage* and *authorial affordances*. Authorial leverage is a concept that originates in a 2009 paper by Sherol Chen and collaborators, and it refers to the ease with which procedural content may be authored in a given AI architecture [chen2009evaluating]. More specifically, when authorial leverage is higher, each unit of authoring effort counts for more. Quoting from an unpublished manuscript, Chen's collaborator Mark Nelson defined it in this way in a recent personal communication:

> The ‘authorial leverage' of a system is the effort-multiplier the system gives the author, i.e. the ratio between the effort it takes to author within a particular AI system and the effort it would take to achieve the same desired effect ‘manually'. [Note: Personal communication, May 4, 2018.]

 In the case of a system with concrete authored characters, retargeting provides authorial leverage by making authoring effort apply to multiple characters at once, as Mateas and McCoy argue [mateas2013architecture, p. 516]. When a system takes the roguelike approach to character authoring (unique ones are generated for each storyworld), retargeting actually becomes a necessity. Without it, an author would have to take the drastic means of authoring procedural content for every possible character, but that is a ridiculous prospect.

The notion of authorial affordances is rooted in the work of Michael Mateas [mateas2003expressive], who defines them as “the ‘hooks' that an architecture provides for an artist to inscribe their authorial intention in the machine” [mateas2002interactive, p. 125]. Retargeting provides such a hook by allowing an author to reason about abstract patterns of behavior that are independent of any concrete characters. This affordance is critical to sculpting the massive possibility spaces of emergent narrative, since it facilitates both the defining and the understanding of such spaces. When it comes to inscribing one's intent into a system, it is critical that the author can not only express her intent, but also understand how the authored material reifies that intent. Abstraction makes authoring more feasible because it reduces the combinatorial authorial burden that would be incurred by enumerating all possible concrete content configurations. In turn, abstraction allows an author to reason more effectively about a system, because it reduces the massive conceptual space of possible content configurations into a smaller abstracted space.

### Modeling Gaps

Simulations have gaps in terms of what they model. That is, all simulations fail to perfectly model the thing they are simulating—this is core to the very definition of simulation, and were the case otherwise for a given system, it would not be a simulation but rather the thing itself. [Note: I should note that it is possible to build computer programs that engage the aesthetic practice of simulation without actually modeling a subject phenomenon in the real world—Conway's *Game of Life* [gardner1970mathematical] is a classic example—but in the case of emergent narrative a simulated storyworld will likely model real phenomena to some degree.] Often, what is modeled in a simulation may call attention to what is not modeled but could reasonably be expected to be, given what is already included. I call these *modeling gaps*.

#### Agency Issues as Modeling Gaps

In a sense, this idea relates quite closely to the notion of *agency* as defined by Wardrip-Fruin and colleagues [wardrip2009agency, p. 1]:

> a phenomenon, involving both the game and the player, one that occurs when the actions players desire are among those they can take (and vice versa) as supported by an underlying computational model.

 In an earlier paper on this sense of agency, Michael Mateas argues that *Quake* is an example of a game with high agency [mateas2001preliminary]. This may seem counterintuitive, because in *Quake* players cannot do much besides moving around and shooting at things. Critically, however, the gameworld as it is modeled does not suggest the possibility of taking other actions; no player is shocked to discover that a Death Knight cannot be engaged in conversation, for instance. [Note: This would make for an interesting mod, though, for that very reason.] As such, this articulation of agency is about a kind of modeling gap, specifically one that pertains specifically to the affordances that are available to an interactor in an interactive work. The connection to the kind of modeling gap that I am outlining here relies on the stipulation that a gap may only be identified relative to what is actually modeled. Thus, while the very ontology of simulation depends on modeling gaps, the stuff that is modeled may call particular attention to what is not modeled. [Note: Note that modeling gaps are in the eye of the beholder. That is, the degree to which a gap is noticeable will vary across individuals, since cultural contexts and other considerations frame the interpretation of a simulation's modeling.]

#### Klein's Modeling Gap: No Detection

An example of this phenomenon occurs in the murder-mystery generator developed by Sheldon Klein and his students. Let us consider an excerpt from “Murder Mystery 1”, one of the outputs generated by the system. This segment occurs after the murder has been discovered:

> The cops arrived. The cops were idiotic. A detective examined the corpse. The policemen looked for hints in the bathroom. Dr. Bartholomew Hume also looked. Edward tried to calm Marion. The policemen questions Dr. Bartholomew Hume. The detective asked questions. The policemen searched the garden. The policemen tried to find clues. Marion cried. Dr. Bartholomew Hume searched stairs. Hume looked for hints. Dr. Hume questioned Lady Buxley. Dr. Hume knew that Lady Buxley told the truth. Florence talked with Heather about the murder. Marion cried. The policemen questioned Ronald. The inspector suspected Ronald. The inspector asked the stupid questions. The policeman searched the parlor. The policemen tried to find hints. Florence was upset. Dr. Bartholomew Hume searched the dining room. Dr. Bartholomew Hume looked for hints. The cops questioned Heather. The detective asked the stupid questions. Dr. Hume questioned Heather. Dr. Hume knew that Heather told the truth. The cops searched the tennis court. Clive talked with Ronald about the murder. The butler said James was kind. The cook talked about the murder. Dr. Bartholomew Hume searched the bathroom. Dr. Hume looked for clues. Marion cried. Dr. Hume questioned Florence. Hume knew that Florence told the truth. Dr. Bartholomew Hume got information from Florence. The cops searched the bathroom. The cops found a thread. The thread was a misleading clue. Lady Buxley talked with John about the murder. Lady Buxley said that James was kind. Dr. Hume was upset. Dr. Bartholomew Hume searched the library. The cops questioned John Buxley. The detective asked the stupid questions. Hume questioned the cook. Dr. Bartholomew Hume knew that Maggie told the truth. Hume got information from the cook. Hume went to the bathroom. Dr. Hume found the bottle. Hume knew the murderer. Hume asked everyone to go to the parlor. Dr. Bartholomew Hume said that the murderer was in the room. Everyone was surprised. Everyone talked. Dr. Bartholomew Hume said that James was killed by poison. Hume said that the butler killed James. Everyone was shocked. [klein1973automatic, pp. 104–107]

This system models the genre of detective fiction, and in particular the murder mystery. One trope in this genre that is clearly targeted by Klein's system is one in which a lay character, amid a cast of bumbling law enforcement personnel, solves the crime through clever deduction. While Dr. Bartholomew Hume does appear to do some detection—he searches rooms and questions some of the other characters—there is no running thread that connects these actions into a process of deduction. A hallmark of detective fiction is the detective's deductive process that culminates in a detailed account of the circumstances of the crime, but it is missing here. [Note: As Ian Horswill noted in his feedback on an earlier draft of this thesis, this feature is less activated in certain traditions, namely the American *hardboiled* genre [chandler2002simple]. Likewise, in his own comments on the draft, Noah Wardrip-Fruin mentioned an alternative pattern that has been called the *reverse whodunnit* [tvtropesreversewhodunnit]: the audience knows all the details upfront, but suspense obtains through uncertainty as to whether the detective will be able to reach the proper solution. Nonetheless, I think Ian and Noah would agree that Klein's system misses the mark.] Because the system targets detective fiction, and because a number of other tropes of that genre are actually modeled—countryside manor, trysts, a murder, multiple character with motives, bumbling law enforcement, amateur detective, etc.—the lack of a deduction simulation is a modeling gap.

In what it is currently to my knowledge the earliest dissertation in all of computational narrative, *Foundations of a Computable Theory of Narrative* (1973), Ed Kahn provides an extensive critique of Klein's system. [Note: Note that Kahn analyzes “Murder Mystery 1” but calls it “Sheldon's Revenge”: “In the story given in Appendix 1 which I have taken the liberty of referring to as “Sheldon's Revenge,” “oversexed Marion” establishes several liaisons.” [kahn1973foundations, pp. 14–15]. This was perhaps meant playfully, but in a phone call last year, Kahn told me that he regretted doing this, especially since it upset Klein (personal communication, June 12, 2017).] As a core part of this critique, Kahn identifies the modeling gap under discussion as the system's fundamental flaw:

> The essential thing lacking in a story such as “Sheldon's Revenge” is any very substantial realization of the detective function. It is not a detective story because there is no detection. Dr. Hume announces that he knows the killer, but never says how this conclusion was reached. There should be an episode of explanation in which the relevant clues are presented. [kahn1973foundations, p. 18]

 Klein, believing that aspects of Kahn's critique were rooted in a misunderstanding of the project, initiated a public dialogue that was published in the April 1974 SIGART Newsletter under the heading “Automatic Novel Writing—An Exchange.” [Note: The crux of Klein's rebuttal is that Kahn conflated Klein's architecture with the procedural data that was authored for it (as a model of the murder-mystery domain). That is, what were in fact expressions of deficiencies in this authored data were misapprehended by Kahn as being expressions of fundamental limitations in the architecture. In fact, Klein's architecture, which was later christened the *Meta-Symbolic Simulation System* (*Messy*) [appelbaum1976meta], is in my estimation an extremely powerful (and elegantly designed) simulation engine. To name a few examples that are familiar to me, this conflation would be like attributing deficiencies in a particular *Tracery*-fueled *Twitter bot* [lampi2017looking] to *Tracery* itself [compton2015tracery], or those of *Prom Week* to *Comme il Faut* [mccoy2014social], or those of *Facade* to *ABL* [mateas2005structuring]. I believe that this phenomenon contributes to the misunderstanding of all sorts of computational systems, particularly generative ones in which the architecture and domain model are decoupled. I propose to call it the *Messy effect*, following a related phenomenon of human misunderstanding, Noah Wardrip-Fruin's *Tale-Spin effect* [wardrip2007three,wardrip2009expressive].] In the exchange, Kahn characterizes the problem of the deduction modeling gap in this way:

> Stories such as MM1 fail to be detective stories because there is no real detection. There is no purposeful acquisition of knowledge on the part of Dr. Hume, he simply announces that the butler, Clive, did it. Real detective fiction involves a step by step enactment of the deduction process. Instead we get a plain announcement, based upon no explicit argument whatever. This failure of be detective fiction is syntactic, the necessary episodes leading to capture are missing. [kahn1974automatic, p. 3]

 In his dissertation, Kahn proposes an interesting solution to this modeling gap, in the form of a detective simulation that would utilize Klein's notion of a “private semantic universe”, which is a mental simulation by which a character recursively runs the storyworld simulation in her mind (with an initial state corresponding to her subjective beliefs about the world) [klein1973automatic, pp. 29–30]. [Note: Klein's description of this technique is quite intriguing. The implementation plan allows a character to literally run a recursive instance of the simulation code, which means that a storyworld (and its evolution over time) truly obtains in the character's mind. For one, Klein notes, this could allow for planning by subjective look-ahead: “An individual character could be made to resort to his own look-ahead simulation of events in order to evaluate decision making criteria about the implication of current actions on future events” [klein1973automatic, p. 19]. More evocatively still, Klein also suggests that the approach could be used to implement character dreams, theory of mind (by running a recursive simulation of another character's mind, given beliefs about their beliefs), and time travel. As Noah Wardrip-Fruin noted in his feedback on an earlier draft of this thesis, Klein's private semantic universes remind him of the way in which *Tale-Spin* [meehan1976metanovel] characters plan ahead by unraveling an extensive series of possible futures. I will discuss this phenomenon, and Noah's own writing about it [wardrip2009expressive, pp. 147–151], in Section 4.2.6.] Here is Kahn's proposal, as articulated in his dissertation:

> It is worthwhile trying to extend Klein's plot specification rules to see what is needed to fulfill the detective function adequately. If Dr. Hume had access to the deep structure of the narrative universe, i.e. its semantics, he would sooner or later observe that Clive was a poor relative of the wealthy victim. If furthermore Dr. Hume had a private semantic universe in which murder motives could be computed and being a poor relative was one such motive, then Dr. Hume would have reason to suspect Clive. Clive's access to poison would be an inference based on his position as a servant in a large house. These three components, a motive model, a means of gathering information for that model, and knowledge about access to weapons, seem minimal to establish some plausible basis for Dr. Hume's otherwise occult source of knowledge. [kahn1973foundations, pp. 18–19]

 In a refreshing moment of candor that could only appear extraneously to formal academic writing, Klein's rebuttal acknowledges the modeling gap and provides a simple explanation for its existence:

> Kahn is right in saying that, in this particular detective story model, the true detective function is absent. This defect is an artifact of the need to construct an example quickly in time for a conference presentation, and in time for my immediately subsequent visit to Moscow on an exchange program. [kahn1974automatic, p. 4]

#### Are Modeling Gaps Political Statements?

In the 1987 edition of his seminal *Computer Lib/Dream Machines*, Ted Nelson writes the following about computer simulation (under the header “All Simulation is Political”):

> Every simulation program, and thus every simulation, has a *point of view*. Just like a statement in words about the world, it is a *model* of how things are, with its own implicit emphases: it highlights some things, omits others, and always simplifies. [nelson1987computer, p. CL149] [Note: It does not require a huge leap to extrapolate Nelson's point as being about all computer programs, not just simulations. Indeed, even seemingly innocuous programs like database systems can be highly political [holmes2008gay]. However, one kind of program that is sometimes viewed as being in a sense unauthored, and thereby free of bias (since there is no explicit creator who inscribes bias into the program), is a neural network. This is not the case, however, since neural networks evolve (*are* evolved) through the use of training data, which will almost invariably encode biases [yapo2018ethical]—*bias in, bias out*. Moreover, as Michael Mateas argues, building and understanding such a model entails subjective human interpretation: “These approaches still require the interpretation of an observer in order to make sense of the input/output relationships exhibited by the system, to select the primitive categories (features) with which the inputs are structured, and to tell stories about the processes producing the input/output relationships. These stories are essential for thinking through which technical constructions to try next, that is, for simultaneously defining a notion of progress and a collection of incremental technical constructions that make progress according to this notion” [mateas2003semiotic, p. 61]. Fortunately, the ethical and social dimensions of machine learning are the subject of a growing research area that aims to raise public awareness of these issues [bostrom2014ethics,brundage2016artificial,yapo2018ethical,brundage2018malicious].]

 In a recent article, Noah Wardrip-Fruin expands on Nelson's statement:

> all the simulations created with these [simulation authoring environments] embed assumptions about the world that derive from viewpoints—they are political. [...] In an individual simulation we see the politics in the rules and data. [wardrip2015we, p. 108]

 Additionally, recent papers by Peter Mawhorter [mawhorter2017efficiency] and Kate Compton [compton2017little] have expounded on this ideas, as have others with particular application to procedural generation [phillips2016feminism,smith2017what,cook2017ethical].

I think there is some additional nuance, which I would like to discuss here, pertaining to the practical aspects of simulation crafting as a creative practice that operates under constraints that ultimately affect modeling. First, computer simulation is constrained by hardware limitations that create computational memory and speed issues; this makes it difficult to simulate complex phenomena. Thus, even if an author has a particular theory about how a subject phenomenon works, hardware limitations may prevent her from operationalizing that theory in a computer simulation. Under such constraints, the resulting simulation may highlight, omit, and simplify in ways that make the simulation deviate from her theory, and thus the program's point of view is not hers. Sometimes an overly simplified model is the only one that is feasible to build. [Note: Interestingly, computationally simpler models sometimes correspond to the simplifications that undergird belief systems that are viewed as problematic [holmes2008gay]. Perhaps the cognitive simplifications result from a similar kind of computational limitation.]

Further, simulation crafting, like any creative practice, has an associated set of skills that, when not honed in a particular creator, will contribute to limitations in that creator's capacity to express herself in that medium. Thus, if I am not particularly good at simulating, it will be hard to express myself through my simulations, and so my viewpoints may not be adequately captured by them. While we understand this more naturally in other mediums—no one thinks a bad painter who unintentionally botches her depiction of a tree is meaning to express a peculiar idea of what a tree is—the notion of skill as a factor in simulation crafting is not often acknowledged (in my experience). It is hard to adequately model complex phenomena!

Relatedly, while stylistic approaches that might be called ‘impressionist' or ‘expressionist' are in the case of other mediums understood to be just that—*stylistic approaches*—simulations are typically interpreted as attempts at literal representation (i.e., ‘realist' or ‘naturalist' expressions). [Note: While I do not know of examples in the area of world simulation, there are text generators that are driven by what I would deem to be impressionist and expressionist inclinations: respectively, Erik Mueller's *Daydreamer* [mueller1990daydreaming] and Fox Harrell's *Griot*, particularly the latter's piece “The Girl with Skin of Haints and Seraphs” [harrell2007griot].] As an example, Anne-Mette Albrechtslund, writing about gender representation in *The Sims 2*, equates simulation with an “aesthetical realism” [albrechtslund2007gender, p. 1]. It is possible to express through computer simulation without attempting to literally represent a subject, and it is important that such efforts be interpreted accordingly. More troublingly still, simulations are often interpreted not only as attempts at literally representing how a subject is, but as moreover instantiating idealized views of how a subject *should be*. As an example, in her paper that I have just mentioned, Albrechtslund assumes that *The Sims 2* encodes the ideological position of its designers [albrechtslund2007gender]. Some simulations may be created in this way, but this is not by default.

Now, I do not mean to claim that all or even most modeling gaps can be accounted for by the preceding justifications. I am compelled to articulate them, however, because in my personal experience with simulation crafting, I have encountered the corresponding misconceptions frequently. There is one last explanation, however, that I think probably does account for most modeling gaps: it takes a lot of time and energy to craft a simulation, and the work is never done, because true verisimilitude (pertaining to a subject phenomenon) can never be achieved (due to the ontology of simulation). It takes time to model anything in a simulation, and invariably there will be gaps that are never rectified. In the case of Sheldon Klein, we know by his own admission that his system fails to model a detection procedure (on behalf of the character who solves the crime) simply because his development team ran out of time and energy. While modeling gaps can indeed reflect a simulation creator's belief system—what she does not think is true (or important) about the subject phenomenon—they may also result from a simple lack of resources. For each of my own simulations, I maintain to-do lists that have grown to document hundreds of modeling gaps or other problems that I have identified and that I hope to rectify. The simple truth is that I will never be able to fill all these gaps.

Before moving on, I would like to emphasize that a system modeling one thing and failing to model another does not necessarily mean that the former is viewed by its creator as being more important. Sometimes an author plucks the lowest-hanging fruits first, thereby saving the more challenging modeling gaps for future work. In the case of Klein's system, I wonder if there is no detection simulation in part because such a system would have to rely on the modeling of phenomena that precedes detection in a detective story—the establishment of criminal motives, a crime, discovery of the crime—and so it would not have made sense to model detection until the earlier subsystems were implemented. Klein did produce the antecedent systems, clearly, but unfortunately he never got all the way to detection.

Finally,

#### Modeling Gaps, Closure, and Apophenia

Finally, it is worth noting that interactors may be inclined, whether consciously or not, to fill in such modeling gaps as a matter of comprehending a simulation or a story. Indeed, even the densest story does not spell out its every detail, but instead the author assumes much about the reader's ability to connect the dots. [Note: The profound inability of computers to fill in such gaps is the impetus for the branch of artificial intelligence called *commonsense reasoning* [davis2015commonsense].] Writing specifically about the space between panels in comics, Scott McCloud refers to the broader phenomenon as *closure* [mccloud1993understanding, p. 63]. Apparently, this concept, as well as others found in McCloud's brilliant *Understanding Comics* [mccloud1993understanding], served as major inspiration for Will Wright as he was developing one of the classic works of emergent narrative, *The Sims* [frasca2001sims,chaplin2006smartbomb]. In the case of simulation, human observers may be inclined to narrate accounts that postulate linkings between emergent events that were not actually linked in the underlying simulation—this is *apophenia* [steyerl2016sea] at work. In Chapter 10, I articulate a simulationist design pattern that I call *apophenia hacking*, whereby apophenia is deliberately targeted.

### Causality Issues

Let us consider a classic distinction between story and plot, originally proffered by the writer E. M. Forster:

> “The king died, and then the queen died” is a story. “The kind died, and then the queen died of grief” is a plot. [forster1972aspects, p. 86]

 Here, Forster is working to develop particular senses of ‘story' and ‘plot' as, respectively, narrative with explicit temporal relations and narrative that also has explicit causal relations. For our purposes here, however, we might simply consider this as the distinction between story and non-story, as others scholars of computational narrative have done [elson2012modeling, p. 44][p. 23][ryan2015narration].

#### Temporal and Causal Relations

What is missing in Forster's first sentence is a relation between the two events beyond the order in which they occurred, that is, something beyond their *temporal relation*. While temporal relations among events are critical in narrative, such relations are not enough to make a story (their presence is a necessary condition, but not a sufficient one). Fortunately, temporal relations are a given in the kind of forward simulation that we find in emergent narrative—the system, omniscient over its modeled purview, knows the order in which its simulated events transpired—which precludes the potentially challenging task of pinning down this information (a task that is central in narrative forms such as investigative journalism and legal argumentation). But, again, temporal ordering alone is not enough: this is the crux of my coadvisor Michael Mateas's critique of emergent narrative as *just one damn thing after another*.

The real juice in narrative is what separates Forster's second sentence from the first: *causality*. In the second sentence, we encounter a meaningful relation between the two events beyond that the first preceded the second, [Note: This temporal relation, by the way, is expressed by the *discourse marker* ‘then' [schiffrin1988discourse, pp. 248–266]. In my first project at UC Santa Cruz, I attempted to automatically discern the *temporal coherence* of blog stories (clarity of the true temporal relations among events) in part by reasoning about the usage of temporal discourse markers [ryan2016recognizing].] which is that the king's death in fact *caused* the queen's death—put differently, there is a *causal relation* between the two events. This brings us to what I view as the most fundamental pitfall in emergent narrative: a lack of meaningful causality between the emergent events of a simulation.

#### The Dense Causal Structures of Life and Simulation

To be fair, simulations by their nature tend to be cybernetical: state evolves as a function of the previous state. In the case of emergent narrative, this means that character actions are typically taken in accordance with some aspects of the current simulation state. In Klein's system, discussed in Section 4.1.3, we saw this in the form of action preconditions that match against asserted facts in the working memory (which stores all the information constituted in the storyworld). Because actions update this working memory and subsequent actions reference it, we can think of the working memory as a mediator that ensures causal linking between events. That is, the causal mechanism is a chaining between the effects of earlier character actions and the preconditions of later actions.

The problem here is that this results in a kind of *massively distributed causality*: a given simulated event will likely be caused by a very large subset (or even all) of the events that preceded it. This occurs when a given action has preconditions that match the effects of many earlier actions, which means that the cause of the action can only be understood as being attributed to (or distributed across) all of the earlier actions. But this can get out of hand: the result of this kind of mechanism is a densely interconnected lattice that basically expresses that all earlier actions caused all intermediate actions to cause all later actions. [Note: In his comments on an earlier draft of this thesis, Ian Horswill referenced a related phenomenon in robotics: “This is sort of like something that comes up in reactive robotics, where you have a bunch of independent behaviors that get triggered bottom-up. If you aren't careful, it will flit between unrelated actions rather than pursing a coherent course of action. That variously gets referred to as ‘chattering', ‘dithering' or ‘lack of hysteresis'{”} (personal communication, July 28, 2018).]

#### The Narrow Causal Structures of Narrative

While the real world may work like this, stories tend not to. In narrative, we find narrower causal structures, with events that cause at most a few other events, and cases of convergence, where the forward causal flows of multiple events converge to cause a single later event. Topologically, these structures resemble sparse directed acyclic graphs, as opposed to the densely interconnected lattices of realistic (and simulative) causality.

Moreover, narrative causal structures tend to bind according to a specific kind of causal link: *contingency*. As the narratologist Mark Alan Powell explains, it is helpful to think about causality in terms of three kinds of causal relationships:

> To understand the plot of a narrative, it is also important to recognize elements of causality that link events to each other. Causal relationships between events may be subdivided into categories of *possibility*, *probability*, and *contingency*. Only in the latter case can one event actually be said to *cause* another. The first category refers to instances when an event simply makes possible the occurrence of another. Relationships of probability, likewise, are those in which one event makes the occurrence of another more likely. [powell1990narrative, p. 40] [Note: Writing about probability in the context of philosophy in 1912, Abraham Wolf identifies these three kinds of causality in addition to a fourth one: *chance* [wolf1912philosophy, pp. 329–330].]

 When a contingency relation linking two events is identified, the earlier event is understood as explicitly causing the later one. Rarely will an event have been contingently caused by more than one or a few earlier events. This is the kind of causality that predominates in narrative: in topological terms, sparse directed graphs whose edges represent contingency relations between events. [Note: The designer and thinker Bret Victor has argued (on Twitter) that the simplistic nature of causality in narrative actually makes it an inadequate medium for evoking empathy: “Narrative inherently embraces simple causes, personal agency, single perspective, limited context. A medium for empathy won't be stories” [victor2015tweet, n.p.]. My labmate Max Kreminski has associated this stance with the catchphrase “stories considered harmful” (personal communication, July 27, 2018).]

#### Inscrutable Causality in *Prom Week *
Simulation feedback should result in what I view as the holy grail of emergent narrative: *emergent contingency structures*. Here, I refer to the kind of narrow causal structures that one finds in narrative, as opposed to the massively branching structures that characterize how causality actually works in the real world and in simulation. That is, an emergent contingency structure is bound exclusively by contingency relations between events, and such a sequence is an important one if it includes an important event. [Note: Something akin to what Labov calls the *most reportable event*. As noted in Section 4.1.1, this is an event that may form the crux of a story.] A nice example of the distinction between this kind of causal structure and a more realistic one (capturing all three kinds of causality) comes in the form of *Prom Week* [mccoy2011prom,mccoy2013prom].

As I explained in Section 4.1.2, in the game's underlying simulation, characters decide what they want to do by reasoning over the state of the storyworld according to thousands of *social considerations* [mccoy2011prom]. Specifically, each of these considerations is represented by a production rule whose preconditions match against the simulation state and whose effects increment the utility of taking a certain action. So what is the cause of a *Prom Week* character action? It is constituted in all of the rules that fired to increase the utility of that action during the procedure for action selection. More precisely, an action's causality is rooted in all of the *social concerns* captured by all of the preconditions on all of the fired rules. [Note: Here, I am simplifying in a particular way, since characters do not autonomously decide which action to take, but instead determine the handful of actions they would like to take most, which a player then selects between. However, the recipient of an action selected by the player will autonomously decide how to respond to that action; in *Comme il Faut*, this means either *accepting* or *rejecting* the action [mccoy2014social, p. 100].] In each case, the specific causal mechanism that binds such concerns to the selected action is Powell's *probability* relation, since the fired rules work to make an action more likely. Moreover, a whole suite of other executed rules may have worked to *decrement* the utility of other actions that may have been taken instead (had those rules not been triggered), and so these rules are also causes of the selected action (again according to a probability relation).

As such, the causality of a given action in *Prom Week* is distributed over potentially thousands of social concerns. Moreover, this distribution may be fairly even: it may be the case that the social concern that contributed most to the selected action being taken is one that accounted for an extremely small percentage of the total utility computed for that action. This is how causality works in the real world—most relevantly here, the causality driven by social norms—but it is not how it works in narrative, where contingency is most important. Just as characters are exaggerations and simplifications of human beings, causality in narrative exaggerates and simplifies the corresponding real phenomenon.

In the case of *Prom Week*, this extreme richness supports their design goals more than it hinders them, since the aim was to produce a game that would support emergent puzzle solutions according to a kind of *social physics* [mccoy2011prom, p. 235]. Indeed, as Alex Mitchell has noted, through gameplay this underlying richness can be apprehended by the player:

> On repeat readings, the surface behaviour I observed began to reveal the underlying system, and I began to develop an understanding of this system through interaction. [mitchell2015reflective, p. 34]

 As Mitchell explains, this kind of understanding of an underlying system through surface interaction is what Noah Wardrip-Fruin has called the *SimCity effect* [wardrip2009expressive]. [Note: In a recent email exchange, *Prom Week* cocreator Mike Treanor noted that the game's interface does much of the work in enabling this kind of player understanding, primarily through explicit surfacing of some of the hidden mechanisms of the underlying simulation (personal communication, April 23, 2018).] But this understanding must come in the form of a general sense of the simulative mechanisms, since no human could fully comprehend the presence and interplay of several thousand social rules. Indeed, the basis of such understanding must be a capacity to aggregate the fine-grained effects of many rules into a coarser representation of essential contributing factors. Could such reasoning be carried out automatically, as a way of binding character actions to construct an emergent contingency structure? [Note: The inferred action relations would not quite be contingency ones, but rather fine-grained probability relations aggregated into coarser-grained relations that could potentially be understood as *quasicontingencies*. In any event, inferring such relations would yield a more actionable comprehension of the causal forces at work.]

I was not a member of the *Prom Week* development team, but in 2015 I worked on a collaboration that sought to automatically generate narrative summaries of gameplay that would be structured as emergent contingency structures [antoun2015generating]. Specifically, we planned for such a generated summary to appear at the end of a playthrough as a diary entry written from the perspective of a character who had been central during gameplay. It would be a generated story recounting what had happened during the playthrough. To enable such procedural generation, we needed an automatic means for identifying why characters took the actions they took and how those actions related to one another in terms of contingency.

Our general approach to this challenge was to isolate a small set of social rules whose execution led to a given character action being taken. The identification of such a set, we figured, would hinge on a particular numeric value: the difference between the utility computed for the selected action and the utility computed for the runner-up candidate action. [Note: Again, matters are confused by the fact that some character actions are selected by players from a pool of high-scoring candidates. It is not worth dealing with this subtlety here.] Let us call this utility difference $x$. So, for example, if at some decision point the utilities of the two highest-scoring candidate actions were computed at ‘122‘ and ‘81‘, respectively, then $x$ would be ‘41‘. Having determined $x$, a system could then identify a set of executed social rules whose utility valences summed to at least $x$. This set of executed rules would be meaningful because, had they not all fired, the other candidate action would have been selected instead. Moreover, such a set of rules could be utilized to automatically generate a causal explanation for the action (one that would express the social considerations encoded in the preconditions of those rules). [Note: Preconditions operating over the *social facts database* [mccoy2010authoring, p. 5] would be especially useful, since entries in that database represent past events. This could allow for explanations like ‘Zack decided to ask out Monica because Monica had done something funny recently', or more complex examples such as ‘Doug insulted Oswald because Oswald recently bullied Monica and Doug considers Monica a friend'. We also considered the prospect of including negative utility valances in this process. For this, we could consider a set of negative utility valances for the second highest-scoring action whose absolute values sum to at least $x$. A set of rules retrieved in this way would capture that the selected action would not have been taken if this set of fired rules had not decreased the utility of the runner-up (to make it less appealing than the selected action); this is an example of *counterfactual reasoning* [ortiz1999explanatory]. A hybrid approach is also possible, meaning one that combines both rules that added to the utility of the selected action and rules that subtracted from the utility of the runner-up.] This procedure would be an example of *abductive reasoning* [aliseda2006abductive], where explanations are generated to account for observations.

While this seemed like a good approach, it did not work in practice. The problem was that such a rule set might comprise several dozen rules, but that would make the resulting causal explanation too rich for human comprehensibility. Next, we considered isolating only the top three contributing rules, but in many cases these rules, even when taken together, could not produce a compelling causal explanation for an action (since they may have only captured a small percentage of the total utility computed for that action). In the end, we more or less abandoned our goal of reasoning over utilities to infer causality, because the utility computation in *Prom Week* was too rich to provide causal accounts that would be comprehensible to humans. Since it did not seem possible to discern the cause of even a single action, the prospect of inferring something on the order of an emergent contingency structure was out of reach. We could not tell stories about *Prom Week* gameplay.

#### Contingent Unlocking in Klein's System

How can a simulation facilitate the kind of reasoning that is needed to discern emergent contingency structures? It turns out that Klein's system, as early as it was developed, provides an elegant solution.

One special kind of action effect in this system that I did not discuss in Section 4.1.3 triggers updates to a critical set of *story registers*. These are data stores that record information about emergent events as a way to guide the generation of subsequent emergent events. For example, during the daytime, all the characters engage in various social activities at Lady Buxley's countryside manor—tennis, chess, conversation, drinking together—and over the course of such an activity two characters may flirt with one another (depending on their personalities, affinities, and other factors). When two characters emergently flirt while being observed by another character, the system records this information by adding a ‘(flirter, flirter, observer)‘ triple to a special flirtation story register.

Later on, either during the daytime activities or at night after everyone has gone to bed, a special follow-up event is unlocked: the flirters may run off together for a secret tryst that is witnessed by the original observer. Once this happens, a suite of additional emergent follow-ups may be unlocked, depending on the relationship between the observer and the lovers. For example, if the observer's spouse is among the lovers, the observer may acquire a motive for killing either the spouse or the other person. Alternatively, the observer may threaten one or both of them for the same reason, in which case the threatened parties may obtain a motive for killing the observer. In the case that a spouse is not involved in the tryst, the observer may blackmail the other parties, who will thereby obtain a motive for killing the observer. Just as a flirtation register tracks the characters who qualify for having a tryst, a motive register compiles all the characters who have unlocked the affordance of murdering another character. Finally, the murder scene occurs emergently when a character from the motive register attacks one of the characters for whom there is a motive to kill.

{Thus, in this way, emergent actions in Klein's system *unlock* future emergent actions that may build upon them to grow full-fledged emergent storylines—such unlocking is an expression of the contingency relationship (causality with a simple and legible cause), and thus such emergent storylines are precisely what I have termed ‘emergent contingency structures'. I call this method *contingent unlocking*. For me, the phenomenon suggests the image of a plant growing out of a spongy forest bed: meaning is constituted against the soft backdrop of entropy.} [Note: In an email discussing prospects for collaboration, Angus Forbes has proposed visualizing emergent narrative using this explicit representational metaphor (personal communication, November 5, 2017). I think that is an exciting idea.]

#### Causal Bookkeeping

While the kind of causality at play in *Prom Week* is a latent one—actions change the world, and the state of the world determines subsequent actions—causality in Klein's system is *explicitly* modeled by the system in terms of contingency relations. [Note: To be clear, there are visualizations of *Prom Week*'s playtrace data [mccoy2014social, p. 108] that resemble the graphical structure of Klein's system's narrative possibility space as depicted in Figure 4.3, but these are different in kind: rather than causal event sequences, these illustrations of playtrace data represent *temporal* sequences (actions that players took).] Beautifully, this allows the system itself to also *identify* the explicit causal relations that bind actions. Thus, by letting emergent actions explicitly unlock subsequent emergent actions—again, I call this ‘contingent unlocking'—emergent contingency structures can not only be generated, but also be easily identified later on. [Note: The task of identifying such sequences in the larger morass of simulated material is *story sifting*, and it forms the backbone of my framework for curationist emergent narrative. Story sifting is explained in Section 5.3.1.] To facilitate such identification, however, the system should explicitly record contingency relationships between events as they occur—I call this technique *causal bookkeeping*. As I discuss in Part II of this dissertation, contingent unlocking and causal bookkeeping have become the most important tools in my technical practice. [Note: After five years of thinking and coding, I independently converged on a similar approach to Klein's. It was not until I read Marie-Laure Ryan's lucid account of his system [ryan1987heuristics] that I realized he had essentially already solved the problem. I think this kind of technique is one that is so elegant and clearly well-founded as to seem obvious once it has been noticed.] Unfortunately, Klein's generated prose does a poor job of rendering the emergent contingency structures that are enabled by its contingent unlocking, in part because it does not do causal bookkeeping in tandem: while emergent events contingently unlock subsequent ones, the system does not record this information for exploitation at the level of curation. Later, in Section 4.2.3, I discuss how doing so would have allowed the system to produce far better narration of its murder mysteries.

So, the takeaway for this section, and perhaps this entire chapter, is that simulations that do contingent unlocking and causal bookkeeping will be far more likely to show not only a capacity for generating satisfying emergent storylines, but for identifying (and ultimately surfacing) them as well. As such, these techniques may be thought of as underpinning a regimen for good *causal hygiene* in character simulation. Like many of the ideas reported in this dissertation, my thinking on this subject was first presented in a 2015 paper called “Open Design Challenges for Interactive Emergent Narrative” [ryan2015open].

#### Contingent Unlocking in *Prom Week *
Briefly, before concluding this section, I would like to note that *Prom Week* actually *does* utilize contingent unlocking, but it was introduced toward the end of the game's development and does not feature heavily in the final product (or its scholarly reporting). This mechanism comes in the form of *story sequences*, which are emergent sequences that play out in a series of explicitly connected *instantiations*. In *Prom Week*, once the player has selected an action for a character to take with another character, the system decides how the receiving character will respond, before finally rendering a concrete performance of the action by the two characters [mccoy2013prom, p. 4]. These realizations, or ‘instantiations', couple generated natural language and character animation. Story sequences, then, emerge at the level of a series of instantiations, where explicit and sometimes evocative connections are made to connect the individual instantiations. For example, one story sequence follows the arc of *A Christmas Carol* [dickens2003christmas], with instantiations concerning visits from Jacob Marley, the Ghost of Christmas Past, the Ghost of Christmas Present, and the Ghost of Christmas Yet to Come, as well as a conclusion scene in which a *Prom Week* character realizes the error of her ways. [Note: Personal communication with Ben Samuel, April 30, 2018. Though for its developers it was a beloved element of the game, *Prom Week* lead author Aaron Reed once told me that he was not sure that any player had ever actually experienced this remarkable story sequence. This was because it was authored in such a way that it could only emerge out of very specific gameplay contexts, which suggests a pitfall that is germane to all computational media: authoring effort may be exerted to produce content that is rarely or never encountered in experiences with a work. Indeed, Ben Samuel recently informed me that he authored no less than fifteen instantiations for this particular story sequence (personal communication, April 30, 2018).]

I do not believe the notion of story sequences has been discussed in any of the papers on *Prom Week* or its technical substrate, but a few years back cocreator Ben Samuel gave me an overview via email:

> Story Sequences was an idea that we got really excited about, but either through our execution, our presentation, or a combination of the two (or something else entirely), they were never a particularly well advertised part of the system. A story sequence, essentially, was a mini-story arc; a story that spanned multiple instantiations. [...] But how could you do this in Prom
Week? With all of the different instantiations, how could you guarantee that the first instantiation in the sequence came before the second? Why, with [labels in the social facts database], of course! [Note: Personal communication, March 8, 2014. My explanation of story sequences relies on the information provided in this email.]

 In *Prom Week*, a special data store called the *social facts database* (SFDB) files records that index simulated events according to a set of author-defined social labels [mccoy2014social, p. 101]. For instance, if one character insults another, the system might file the insult under the label ‘rude‘.

To support story sequences, the authors invented a hack of this labeling mechanism by which the component actions of a sequence are filed under special sequence-specific labels (e.g., ‘SFDBLabel_SS8_1‘ for the first action in the eighth authored story sequence). In turn, later actions in the sequence will have preconditions that demand the presence of such sequence labels in the social facts database (so that, for example, the second instantiation in a story sequence does not appear out of context). Moreover, to increase the likelihood of story sequences emerging, the game's creators authored rules that would boost the computed utility of the components of story sequences.

It should be clear by now that this SFDB hack bears a striking resemblance to Klein's story registers. While the system does not explicitly do causal bookkeeping, it could be trivially implemented by recording cases of the special SFDB labels unlocking the next instantiation in the story sequence. Unfortunately, as Samuel noted in his email to me, *Prom Week*'s story sequences did not end up becoming a major part of the gameplay experience or the scholarly reporting on the system. Nonetheless, the subsystem is a nice example of contingent unlocking in emergent narrative.

So while Klein's murder-mystery generator and *Prom Week* both make use of contingent unlocking to support emergent contingency structures—what I believe to be the holy grail of emergent narrative—no one seems to have noticed. As with many of the other pitfalls outlined so far, this appears to be due to the relatable case of developers who could not quite close the loop on execution (or reporting) due to a lack of time, energy, or other resources. Moreover, neither system explicitly carries out causal bookkeeping, but in either case it would have been feasible, as I have shown. Nonetheless, each project's approach to contingent unlocking serves as a good example of the kind of technical approaches that can be utilized by practitioners building their own systems.

## Curation Pains

While simulation issues may hamper the quality of stories produced in works of emergent narrative, curation issues trouble the very classification of those artifacts as stories. In this section, I begin with what I have identified to be the essential flaw of emergent narrative—the mere transpiring of simulation is conflated with narrative—before proceeding to articulate a series of additional curation issues that have worked to hamper the form over the course of its entire history.

### No Telling

There is a troublesome conceit in emergent narrative that the mere transpiring of a simulation, or of an interaction with one, constitutes narrative. This notion is especially prevalent in interactive works, including games such as *The Sims* [maxis2000sims] and academic projects like *FearNot!* [aylett2005fearnot]. My view, as I alluded to above and explain more thoroughly below, is that a narrative artifact only manifests through actual curation of the simulation stream. That is: *raw simulation is the stuff of emergent narrative, not the shape of emergent narrative*. I am not the first person to offer this kind of critique, and in this section I will make heavy use of specific arguments by earlier writers.

#### The Raw Unfolding of Time is Not a Story

As Espen Aarseth notes in his seminal *Cybertext* (1997), the idea that simulation is a kind of narrative is likely rooted to some degree in the *narrative turn* [kreiswirth2005narrative,czarniawska2004narratives] (or narrative turn*s* [hyvarinen2010revisiting]) in the humanities and social sciences:

> There is a tendency in much cultural theory to posit narrative as the grand structure of everything, the foundation upon which we order our lives and actions. To suggest that narrative is not wholly deserving of this reverence might be risky, since it is all too easy to point out that even the very point I am making here could not be made without the support of narrative. [aarseth1997cybertext, p. 94] [Note: Later, building off of Markku Eskelinen [eskelinen2001towards], Aarseth would also identify a “narrativistic colonialism” by which “the (academic) discovery of computer games over the last two decades is accompanied by the most smothering form of generic criticism: the attempt to reform games into a more acceptable form of art, literature or film; i.e., as narratives” [aarseth2004genre, n.p.]. Scholars such as Jesper Juul backed this view [juul2011half, p. 16], while others took offense at the phrasing [murray2005last,simons2007narrative].]

I too am reliant on a kind narrative mode in this very writing, and I am in fact someone who subscribes to the ideas that drive the narrative turn, but it is critical to realize a subtle distinction: narrativist scholars and researchers do not posit narrative as the basis of human experience, but rather as the basis of *human sensemaking*. {When narrative psychologists write of the “storied nature of human conduct” [sarbin1986narrative], they do not mean that human conduct itself constitutes narrative—that the raw unfolding of time is literally a story—but rather that *we can only understand it as a kind story*, since we process the world and our experiences in it through a mechanism of narrative sensemaking.} As Theodore R. Sarbin makes clear [sarbin1986root], narrative by this perspective is a *root metaphor* for human conduct. The notion of a root metaphor is due to the philosopher Stephen C. Pepper [pepper1942world], who defines it as a fundamental organizing principal for sensemaking (both formal and informal):

> By a root metaphor, I mean an area of empirical observation which is the point of origin for a world hypothesis. When anyone has a problem before him and is at a loss how to handle it, he looks about in his available experience for some analogy that might suggest a solution. This suggestive analogy gives rise to an hypothesis which he can apply towards the solution. The method of development of world hypotheses for the problem of gaining comprehension of our world, follows, I find, the same procedure. [pepper1966concept, p. 3]

 Thus, narrative by this view is a mechanism that can be applied to make sense of the world, *and that application works metaphorically*, since the raw material of the world is treated as narrative *by metaphor*. But, again, this does not mean that the raw material is literally narrative.

While narrative psychologists (and humanists and social scientists and philosophers and other professional thinkers) have applied narrative sensemaking to make sense of their subject phenomena, they also describe how all human beings carry out this same procedure to make sense of everyday experience. In this way, as the basic mechanism of human understanding, narrative sensemaking drives conception of all human conduct, and moreover conception of all conduct whatsoever, inasmuch as human understanding obtains. In fact, this sensemaking is actually *overactive*—studies have found that humans may even perceive random event sequences as constituting narrative [sarbin1986root]. [Note: Richard Walsh provides an interesting temporally oriented account of the human tendency to narrate: “Narrative, after all, is our principal way of understanding our experience of reality by articulating patterns in time; if this cognitive strategy misleads us in relation to emergent phenomena, it may be something to do with the fact that narrative itself—as instance, as mode, and as cognitive faculty—is itself an emergent pattern in time” [walsh2011emergent, p. 84].] Still, while essentially any happening may be understood as constituting a kind of narrative (or at least eliciting narration), this does mean that the happening is literally a story. To maintain as much is to conflate metaphor with an equivalence relation. [Note: If all phenomena are understood through narrative sensemaking, then stories themselves must also be processed in this way, which means that humans mentally construct (meta)stories to make sense of stories. Does this then mean that there are no stories constituted in the external world, but instead only stories that humans tell themselves about those stories? While the subjective nature of story comprehension is likely unavoidable [bakhtin1981dialogic,duranti1986audience,goodwin1984notes], this line of thinking is probably not that constructive here. By assassinating objectivity, as we did in Section 3.1.1, we can destroy all delineation and demarcation, but then we are not left with much to talk about. Just as there is a thing called ‘nonfiction', there are things called ‘stories', and so for the purposes of this dissertation I will assume that ‘stories' are in fact stories. (A thought: if humans must construct metastories to make sense of stories, then it follows that in turn humans must construct metametastories to make sense of the metastories, and so forth to some limit of recursion rooted in either human understanding, human metaunderstanding, brain physiology, or the material universe.)]

#### The Raw Unfolding of Simulation is Not a Story

Now, we may return to Espen Aarseth, who is likely the scholar most associated with the argument that raw simulation is not literally narrative. On the same page in *Cybertext*, Aarseth continues:

> But the story of an event is not necessarily the same as the event itself, and stories can be told about things other than stories, luckily. [...] For instance, both stories and games of football consist of a succession of events. But even though stories might be told about it, a football match is not in itself a story. The actions within the game are not narrative actions. [aarseth1997cybertext, p. 94] [Note: Incidentally, as noted in Section 2.2, in her first articulation of the approach, Ruth Aylett explains emergent narrative by making an analogy to the ways in which “recognizable narrative structure” emerges in a soccer match [aylett1999narrative, p. 84].]

 In his own invocation of this sports analogy, the scholar Markku Eskelinen did not mince words:

> if you actually know your narrative theory [...] you won't argue that games are (interactive or procedural) narratives or anything even remotely similar. Luckily, outside theory, people are usually excellent at distinguishing between narrative situations and gaming situations: if I throw a ball at you, I don't expect you to drop it and wait until it starts telling stories. [eskelinen2001towards, pp. 175–176]

 Later, in a 2004 paper, Aarseth argues that stories made up of simulated events are typically constructed *after the fact*, by the process that I call curation:

> Nevertheless, games like *The Sims* are sometimes (not often) used as storytelling machines, when particularly memorable moments in the game are retold by the player/god. But this is not translation from game to story, this is simply good old after-the-fact narration, like the football column in the Monday sports section, the lab experiment report, or the slide show of one's Carribean vacation. Something interesting happened, and we want to tell others about it. Ontologically, the capacity for generating memorable moments is something games have in common with real life, as well as with stories. A story-generating system does not have to be a story itself. In fact, while life and games are primary, real-time phenomena, consisting of real or virtual events, stories are secondary phenomena, a revision of the primary event, or a revision of a revision, etc. [aarseth2004genre, n.p.]

 In the interim between these papers—and in the midst of the *ludology vs. narratology debate* in game studies [frasca2003ludologists,murray2005last,mateas2005build], which also structures Aarseth's post-*Cybertext* writing—Jesper Juul argues that there will tend to be (or there necessarily will be) fundamental divergence between gameplay as experienced and gameplay as recounted:

> if we recount a game of chess, our playing of the entire Half-Life game or a multi player game of Starcraft, the existents and events will be transferred, but not the dynamic systems. Our retelling will not be a game, and in fact much of the vast journey that it takes to complete Half-life would be excruciatingly dull if retold in any detail. [juul2001games, n.p]

 Finally, in a more recent 2012 paper, Aarseth pains himself one last time (for now) to consider the view under fire in this section, which he terms *narrativism*:

> This conflation of any kind of diegetic or experienced situation with storytelling is what I have
previously labeled narrativism [...] If an(y) interesting experience in a game is an “emergent narrative,” where does it end? And why limit this category to game-based situations? At some point it becomes hard to distinguish narratives from any other type of worldly experience, at which time (or long before) we might as well give up the discussion. [aarseth2012narrative, p. 131] [Note: In his account of the ludology vs. narratology debate, Frasca uses the term ‘narrativism' to refer to the stance taken by the contingent conventionally called the ‘narratologists' [frasca2003ludologists]. (He does this, following Michael Mateas [mateas2002interactive, p. 32], from the belief that ‘narratology' was being used incorrectly in the debate [frasca2003ludologists, p. 2].) This does not appear to be the sense with which Aarseth uses the term. Instead, as he has noted [aarseth2004genre, n.p.], he is invoking what Alan Rauch has termed “story fetishism” (extremism in the narrative turn outlined above). Curiously, my attempts to track Rauch's usage of that phrase only turn up Aarseth's quotation.]

While Aarseth and his peers are specifically concerned with videogames and interactivity, I think they hold just as well for systems outside videogames (such as story generators) and for noninteractive cases. Indeed, the narrative theorist Richard Walsh argues that, in any case, emergent narrative obtains because we are inclined to narrativize (alternatively: curate):

> a particular run through a simulation [...] does not in itself produce emergent narrative but instead emergent behavior. So Aylett's virtual Teletubbies, interacting with each other and with the user's input will (given the right conditions) produce emergent behavior that we as observers may be inclined to narrativize. Similarly, the goal-oriented user in [an] interactive environment participates in the emergent behavior that system provides for, and that *experience* is one that the user may be inclined to narrativize. [walsh2011emergent, p. 77]

 Indeed, this misconception seems to be fundamentally rooted in the potential *tellability* of simulated events. As I explained in Section 4.1.1, in the study of personal narrative, events are deemed tellable if they would make for a good story (or a story at all) [norrick2005dark]. It would be strange, however, to argue that the tellable events of lived experience literally constitute a story—the story, of course, materializes in the actual telling. But this idea is what drives the troubling argument that the mere transpiring of simulated events constitutes a story. According to this perspective, a simulated event, *as it transpires*, constitutes a segment in an unfolding story. By analogy, a tellable event in the real world, *as it transpires*, would constitute a segment in an unfolding personal narrative. This formulation is strange because it mixes up what is being recounted with the actual recounting. The unfolding of time itself is not a story, and so neither then is the raw unfolding of a simulation (or some record of that unfolding, such as a simulation trace). Put another way: simulated material is the *stuff* that composes emergent narrative, but it is not itself emergent narrative. To contend otherwise is to conflate simulation with narration.

#### But Simulation Does Afford Narration

There is still something powerful happening here. While tellable events are not a telling, tellability critically *affords* telling—in turn, while simulated events are not stories, they may afford stories. Thus, the fact that simulated events may be tellable at all has a major implication: stories may be told *about* simulations, which means the very *project* of emergent narrative is viable. This project, however, is predicated on a procedure of curation.

Let us briefly recall Hayden White's model of historiography, which I outlined in Section 3.1.1. White argues that an historical account may only materialize through an act of curation, by which the raw material of transpired phenomena (White's ‘historical field') is processed and then embedded into the structural framework of a story. This process entails several steps: the phenomena transpire, a chronicle (lossily) records them, a subset of the chronicle is selected to be included in a story, and finally the story is fit into a particular ‘emplotment' (e.g., the tragedy or comedy), resulting in an historical account with targeted rhetorical features. Just as historical accounts obtain through curation of an historical field, emergent narrative obtains through curation of a simulation. [Note: Jan Simons has also noticed a common essence spanning historiography, personal narrative, fiction, and emergent narrative: “events, whether historical or fictional, actual or virtual, are not intrinsically narrative or non-narrative but they become stories because someone deems them ‘tellable' and perceives or construes them as causally and chronologically connected” [simons2007narrative, n.p.].] By such a procedure, tellable simulated events may actually be told.

#### Offloading Curation to Interactors

Finally, we arrive at the fundamental flaw of emergent narrative: there is no telling. By implying that their raw simulation streams literally constitute narrative, works in this area offload the burden of curation—by which stories may actually obtain—to human interactors. In this way, they are story generators that do not actually tell stories. [Note: Furthermore, a loose idea of emergent narrative being constituted in simulation may lead to a vague conception of what form stories actually take in a project. Hartmut Koenitz has raised this concern with regard to a project that he views as otherwise being successful (citation inserted for clarity): “From this perspective, McCoy et al.'s otherwise outstanding work on Prom Week is marred by the lack of a clear definition of ‘story.' For example, in one of their papers [mccoy2013prom], the authors variously refer to story as ‘personally meaningful' game experience, as ‘campaign,' as ‘gameplay,' as a ‘collection of levels [...] where the player can take social actions,' or they talk about ‘player agency at the story level' and the ‘player's path through the story.' Each of these descriptions depicts ‘story' as a slightly different concept” [koenitz2017interactive, p. 10].]

To be clear, I think this pattern can work quite well in some cases—*Dwarf Fortress* and *The Sims* cannot be denied—but it requires a good simulation and a human willingness to curate. This human curation can take the form of basic narrative sensemaking (players mentally narrativizing the unfolding simulation), but more powerfully it may result in conventional narrative artifacts that are mounted in full-fledged media experiences. Here, Tim Denee's comic *Oilfurnace* [denee2010oilfurnace], the subject of my case study of Section 3.1.1, is a perfect example. Another is Robin Burkinshaw's *Alice and Kev* [burkinshaw2009alice,evans2011representing], which was produced by curating *The Sims 3* [sims3] gameplay. In his recent dissertation, Eric Murnane provides an extensive analysis of player stories that recount emergent experiences in *The Elder Scrolls V: Skyrim* [murnane2018emergent], a practice that he himself has carried out with the extensive ‘narrative photojournal' *Three Hundred Days in the North* [murnane2017three]. Likewise, Stephanie Boluk and Patrick LeMieux write about the ‘Dwarven Epitaphs' produced by players of *Dwarf Fortress*, among which *Oilfurnace* is a featured example [boluk2013dwarven]. While I view these as powerful examples of curationist emergent narrative, the nucleus of this dissertation, {this pattern of human curation does not seem to be prevalent outside the realm of popular videogames. In the case of non-interactive story generation, and indeed many other interactive systems, humans may not be so inclined to tame the welter. To again invoke my coadvisor Michael Mateas's critique: without curation, emergent narrative is *just one damn thing after another*.}

#### Toward Automatic Curation

What I seek to promote in this dissertation is not just an emphasis on the importance of curation in emergent narrative, but moreover an approach *by which systems themselves may actually tell stories*. This is no small task, since it would entail the *automatic curation* of a simulation in a manner that is (in terms of procedure) akin to Hayden White's historiographic method: the system must sift through a record of what has happened in order to identify interesting material out of which a narrative artifact may be constructed. That is, the system must not only be able to tell stories, but it must also have the capacity to discern tellable event sequences in the first place. Further, as I will argue in Section 4.2.4, curation should not culminate merely in a narrative artifact, but moreover that artifact should be mounted in a full-fledged media experience. In the next chapter, I fully articulate my vision for procedural narrative in the form of my framework for *curationist emergent narrative*. [Note: The essence of my argument here first appeared in two 2015 papers [antoun2015generating,ryan2015open] that introduced a challenge that I refer to as *story sifting* (then called ‘story recognition'): the task of automatically recognizing interesting narrative material embedded in the morass of accumulated simulated material. The latter paper, in particular, may be viewed as the embryo of this dissertation. In Section 5.5, I recount this intellectual development in more detail.]

### No Curation

While works of interactive emergent narrative typically do no telling, simulationist story generators may do telling without curation. Whereas in the former case the emergent narrative is viewed as obtaining through the process of a simulation transpiring over time—with its events washing over the interactor, if you will—in the latter case the evolution of the simulation over time is captured in a *trace*, and that trace is what is (mis)treated as a story.

#### *Tale-Spin 's Curious Model of Creative Writing*
This approach is probably best characterized by early systems like Sheldon Klein's murder-mystery generator [klein1971program,klein1973automatic] and especially *Tale-Spin* [meehan1976metanovel]. As Masoud Yazdani writes (in critique, not defense), creator Jim Meehan's approach to the latter project instantiates a peculiar model of creative writing:

> A story [by Meehan's approach] is an account of goal directed behavior of a set of characters—their interactions with each other and with the world. Story writing is simulating a world in which characters follow a goal directed life, presenting the trace of the simulation to the readers. The readers can be expected to be interested in these stories as they are concerned with problem solving and purposeful interaction themselves. They like to find out about such situations and find these stories interesting. [yazdani1989computational, p. 138]

 To be clear, Yazdani's account is actually a recapitulation of an explicit claim that Meehan himself proffers in his dissertation:

> What happens when the simulator runs may or may not be interesting [...] How do you make it interesting? You fix it in advance. You rig the world so that if people do behave rationally, they'll do some interesting things. So sitting on top of the simulator, if you like, is a program which knows both about stories and about the world model. It models a writer who has something in mind that he wants to tell a story about. [meehan1976metanovel, p. 108]

Indeed, the *Tale-Spin* approach may be criticized for its idea of how stories are made—this is what sparked the turn, beginning with Natalie Dehn's *Author* [dehn1981story] and Yazdani's *Roald* [yazdani1982write], toward an author-centric approach to story generation that more intuitively models the process of human story invention. For our purposes, however, I want to call into question a specific feature of this approach, which is its idea of *how a story looks*.

#### A Simulation Trace Is Not a Story

*Tale-Spin*'s generated tales are essentially prose renditions of traces through its simulation, and I believe this aspect of the project has contributed to its lambasting by humanists and later practitioners. Here is an example excerpt from a system output:

> Tom decided that Wilma might want Tom to give Wilma a worm. Tom wanted to ask Wilma whether Wilma would tell Tom where there were some berries if Tom gave Wilma a worm. Tom wanted to get near Wilma. Tom walked from the chair across a living room down a hall via some stairs down a hall down a hall through a volley across a meadow to the ground by the redwood tree. Tom asked Wilma whether Wilma would tell Tom where there were some berries if Tom gave Wilma a worm. [meehan1976metanovel, p. 232]

 A trace of an entire simulation over time does not make for a story, but rather a recording of a *storyworld's history*. In the case of an interactive experience, an interactor may be willing to process the welter of ongoing simulation to sift out the emergent stories, but this is probably not to be expected in the case of a prose artifact. You likely struggled to finish reading even the short excerpt that I just quoted. In Section 4.2.4, I discuss how presentation issues have contributed to much of the failing of emergent narrative, but presentation only comes at the end of a substantial process of curation. Before Meehan's *Mumble* system generates its stilted prose, the material that will be presented has already been selected for rendering in natural language. The process of deciding what material should be included in the story presentation is an early step in the curation procedure, and one that *Tale-Spin* fails to adequately carry out. By the time that its generated prose does further injustice, the system's tales are already doomed because nearly all of the available simulated material is blindly selected for telling, which makes the outputs more like simulation traces than stories.

#### Simulation Traces as Whitean Chronicles

We can think of the curation procedure of emergent narrative as working like Hayden White's historiographic procedure (discussed above): an historical field (simulation) is captured in a chronicle (trace or other data), out of which a subset of material is selected for use in a story (story) that is embedded into an emplotment (is deliberately crafted) with particular rhetorical features.

When a monolithic simulation trace is printed out, a Whitean *chronicle* is produced, not a *story*. Indeed, Stephanie Boluk and Patrick LeMieux have made essentially this observation with regard to the simulation traces found in *Dwarf Fortress*'s Legends mode [dflegendsmode]:

> *Dwarf Fortress*'s linguistic forms of historical inscription in Legends mode bear a striking formal and thematic resemblance to early forms of writing, such as the medieval annal and chronicle. [boluk2013dwarven, p. 144]

 As I discussed at length in Section 3.1.1, White explains how chronicles have no beginning (they merely start), no end (they merely terminate), and no meaningful structure in between. To tell a story out of external material—this is the essential task of historiography—the teller must carry out a process of deliberate selection and deliberate assembly of the selected material into the frame structure of narrative. To be clear, this means selecting a subset of the material contained in the chronicle. If instead a chronicle is itself presented as a story, the recipient will deem it either dubious or non-narrative—she will likely be unwilling to sift through it to find the story (only historians and lawyers do this). This, I contend, is what drives the harsh critiques of Klein's murder mysteries and *Tale-Spin*'s fables—simulation traces are chronicles, not stories. [Note: Boluk and LeMieux, who also connect simulation traces to Hayden White, do so with favorable appraisal of the former: “By contrast, White argues that the annal and the chronicle are not ideologically primitive forms of inscription but instead embody a fully developed but radically different philosophical worldview. He proposes that these forms of medieval writing have been misrecognized in much the same way that it is easy to misinterpret the output of *Dwarf Fortress*. Like *Dwarf Fortress*, the annal deprivileges the place of the human within a larger cosmological landscape” [boluk2013dwarven, p. 145]. This line of argument gibes with my call, coming in Section 3.1.4, to embrace a new aesthetics of the computational—a kind of *computer art brut*—but I am not sure simulation traces are the right artifact for the job. Often traces seem to be the lazy first approximation of an actual considered artifact. I think we can do full-fledged emergent narrative without losing the evocative aesthetics of the computational, but doing so means carrying out a procedure of curation. This is the primary call of this thesis.] Later, in Chapter 8, I apply this critique to my own work in a takedown of my project *Diol/Diel/Dial*.

#### Klein on Curation

To be clear, earlier practitioners did realize the importance of curation to some degree. Meehan knew that his simulation traces were not good stories, and Klein admitted, “It is a rather long distance from the promise of the title [a novel] to the fragmentary output just presented” [klein1971program, p. 30]. Later, in an interview with *Byte Magazine*, Klein explicitly emphasized the importance of curation with regard to his computer-generated video opera *Revolt in Flatland* [klein1982propositional]:

> Klein and his colleagues showed a video-taped sample at the Fifth International Conference on Computers and the humanities, in Ann Arbor, Michigan, last year. But, says Klein, “I'm not keen on showing it because, to be perfectly frank, the music is superb but the action is quite dull.” This is because the action simulator treats the cast's movements as if they were taking place on a realistic scale, rather than compressed onto a stage. “Most of what's happening in the current version is that little squares and triangles and polygons are moving slowly from one house to another. The action only happens occasionally, when they meet,” Klein says. A refined version of *Revolt in Flatland* is being prepared which automatically omit the boring scenes. In the meantime, says Klein, “Potentially, our opera could be five hours long, like a real opera, but at the end there would be nobody watching.”
[mckean1982computers, p. 51] [Note: Unfortunately, a refined version was never produced. I reached out to Mark Manasse, one of several students who worked on the project with Klein (it was being undertaken for his seminar class), to ask what happened. As he explained, a strike critically halted work on the project (and presumably blocked computer access, as well): “We had a lot of ambition, but the class coincided with a strike by the teaching assistant's union, which greatly interfered with [our] implementation plans [...] to allow our ambitious system to do its work; the design existed on paper, but never got finished” (personal communication, April 20, 2018).]

#### *Saga II 's Brilliant Curation*
Moreover, one of the early projects actually demonstrates an emphasis on curation that would be exceptional even today. In *Saga II* (1960), discussed in Section 4.1.2, there is no conceit that raw simulation traces are narrative artifacts or full-fledged media experiences. Instead, the printouts containing such traces were subjected to additional curation, by which shooting scripts were developed and further curated to drive the production of a television scene. This emphasis on curation—and mounting in a full-fledged media experience, which is the call of Section 4.2.5—was likely due to the project being a collaboration between computer scientists and television producers. Nonetheless, it makes *Saga II* a fabulous example of the power of curation in emergent narrative. In Chapter 5, I celebrate this achievement further.

#### Even Traces Are Curated

Before moving on, it is worth noting that even chronicles and interactive simulations result from a process of curation. Indeed, as I will explain in Section 4.2.6, *Tale-Spin*'s simulation traces are in fact highly curated with regard to the larger accumulation of material that is produced in a given run of the system. [Note: Preview: *Tale-Spin* actually curates *out* all the interesting material.] The same goes for Sheldon Klein's murder-mystery generator. This is actually quite intuitive: when a simulationist method is employed, the system is not operating at the level of natural language, or even plot. Rather, a world is simulated, and then natural language is generated by a separate procedure to express what has happened in that world. Sheldon Klein and Jim Meehan were very explicit about this distinction, which is why Klein presents his system as primarily a contribution to natural language processing, while Meehan apologizes for his hasty *Mumble* hack that was cobbled together in a day.

#### Representation is Curation

What I am getting at here is that a simulation cannot be presented, or even experienced, in its canonical form. A simulation trace is a *report* about a world, clearly, but even the storyworlds of interactive simulations cannot be truly encountered—humans cannot actually experience the text worlds of interactive fiction or *multi-user dungeons* [ito1997virtually,young1996computer,ryan1997interactive], and this is true also of graphical worlds, be they 2D or 3D. Indeed, the very descriptors ‘text world' and ‘graphical world' reveal this: those worlds are not actually made of text, and they are not actually made of pixels, but rather that material is the stuff out of which *representations* of those worlds are fabricated. We may be tempted to say that such representations *render* the underlying worlds, but rendering is actually a mode of curation—a textual or graphical representation of a world, even if it is interactive, is a *depiction*, and depictions are not their subjects themselves.

As Michael Mateas has argued, technical strategies for the implementation of computational systems are often conflated with the stories we tell about those systems—what he calls the “inseparable, tightly entangled collection of rhetorical and narrative strategies for talking about and thus understanding these computational systems” [mateas2003semiotic, p. 61]. One such rhetorical strategy is textual or graphical rendering of an underlying world, the canonical form of which would be incomprehensible to a human. [Note: Here, we converge on fundamental issues at the heart of the philosophy of computation, and moreover, of cognition. What is a *Tale-Spin* world? If it grounds out as data stored in memory, a series of ‘on‘ and ‘off‘ symbols, how do those ground out? Is it in the stored physical charge of the electromechanical system inside the computer? But that stored charge will take a different form on a different computer, even if the world is the same (“the same”), and such divergence may also occur across different runs on the same computer. So maybe the world does actually ground out in a symbolic representation, according to a mapping scheme by which some mediator, be it human or computer, can maintain *isomorphism* for the world across its different physical manifestations. But then we hit *Searle's wall*: “For any program and for any sufficiently complex object, there is some description of the object under which it is implementing the program. Thus for example the wall behind my back is right now implementing the Wordstar program, because there is some pattern of molecule movements that is isomorphic with the formal structure of Wordstar” [searle1992rediscovery, pp. 208–209]. While this thought experiment is more obscure than his *Chinese room* [searle1980minds], it is far more ramifying: “But if the wall is implementing Wordstar, then if it is a big enough wall it is implementing any program, including any program implemented in the brain” [searle1992rediscovery, p. 209]. This is the way toward *panpsychism* (the view that all matter, and each grouping of matter, has a mind [seager2005panpsychism]), but James Blackmon comes to the rescue. As he lucidly explains [blackmon2013searle], mapping schemes are only *frames of reference*—just as any object may be perceived as traveling at any speed according to the properly concocted frames of reference, any (sufficiently big) material system may be perceived as computing anything if the right mapping scheme is concocted. The problem is that it is hard to concoct such schemes—so, yes, while the completed form of this dissertation, as filed, has existed in my wall all along, I had no means (no mapping scheme) with which to discern it. The monkeys have already typed infinite pages, but finding the ones we want means walking in a Borgesian library. In a sense, however, this is what computation is: we build computer programs as mapping schemes that allow us to discern—in the welter of matter all around us, in which everything ever is always already constituted—the particular phenomena that we would like to encounter. We do not create the images, but rather we capture them as they course through the electromechanical network constituted in the machine. The computer is a weathervane. So while all the *Tale-Spin* worlds, mere images, are already all around us all the time, Meehan's software (or Warren Sack's [sack1992micro]) provides the mapping scheme by which we may isolate those images, so that we can encounter them.] To use Espen Aarseth's words for cross-purposes, this conflation seems to be rooted in “a spatiodynamic fallacy where the narrative [world representation] is not perceived as a presentation of a world but rather the world itself” [aarseth1997cybertext, pp. 3–4]. [Note: Here, I am pitting Aarseth's words against him, since he used them to argue for a world essentialism in nonlinear narrative: while a trope in literary studies suggests that the ‘world' of a text may be encountered (and explored) through reading, Aarseth clarifies that this is actually a metaphor. His intended point, however, is that such topology is actually present in nonlinear texts: they present worlds that may truly be encountered. I am taking his argument further, past his point, by contending that it is *still* a metaphor in the case of nonlinear texts, and moreover even in the case of graphical worlds. While they may seem explorable in a phenomenological sense, these media worlds are just images, like the shadows on Plato's cave wall.]

All *representational strategies* [mateas2009defining] for rendering storyworlds are constituted in *rhetorical strategies* [mateas2003expressive], and together these strategies work to compose larger curation procedures by which underlying storyworlds may be encountered by human beings. One possible mode of curation is textual or graphical representation, and another entails the maintenance of an interface through which a textual or graphical world may receive human inputs. At the heart of this document is a further mode, by which the world is curated so that a narrative artifact is constructed for human encounter (and then mounted into a full-fledged media experience). Of course, these are not incompatible modes, and indeed they may be stacked and recombined in interesting ways. My point is that it does not seem possible to avoid curation—so why not take up the call of this dissertation and be deliberate about the process?

#### No Emphasis On Curation

You may have noticed that I have just defeated my own thesis for this section—that simulationist story generators like *Tale-Spin* have failed in part due to a lack of curation—but I contend that the point still stands. While Meehan's system does do curation (it is impossible not to), it does so in a haphazard way that reflects a lack of emphasis on this critical aspect of emergent narrative. As such, a more precise title for this section would be ‘No Emphasis On Curation'. In the next section, I will identify particular ways that curation can go wrong, even in cases where a system builder emphasizes its importance.

### Poor Curation

Since the procedure of curation entails nearly all the same work as conventional narrative construction, a number of pitfalls may befall the (automatic) curator. Generally, it seems that things can go wrong in four ways: poor selection of material, poor arrangement of selected material in the telling, poor presentation of that arranged material, and a failure to mount that presentation in a full-fledged media experience. The preceding section was about an extreme case of poor selection of material (the wholesale selection of a monolithic simulation trace), while this section  pertains to more practical considerations of the curation phases in which simulated material is selected and then arranged for recounting. In the remaining sections of this chapter, I discuss the issues of presentation and mounting in an actual media experience.

#### Klein Kills the Suspense

Again, we find in Sheldon Klein's murder-mystery generator a particularly illustrative example. There is one particular blunder that contributes more than any other to the failings of this system's synthetic tales: the killer is always revealed halfway through the story! As Figure 4.3 illustrates, the murder tends to occur about halfway through each simulation run. But while a delayed revelation of the details of the crime (and particularly the identity of the killer) is the hallmark of the murder-mystery genre, Klein's system divulges these details *as the crime is transpiring*, as this example murder scene shows: [Note: One could argue that Klein's system actually generates *reverse whodunnits* [tvtropesreversewhodunnit]. In this mystery mode, the audience knows all the details of a crime, but suspense still obtains through the mystery of how a detective will solve what appears to be a perfect crime. This is hard to argue in the case of Klein's system, however, since there is no rich modeling of detection, as discussed in Section 4.1.5.]

> Lord Edward knew that Lady Jane commited adultry [sic]. Lord Edward was enraged. Edward decided to stab Jane. The day was Sunday. The time was the sunrise. Jane awakened early. Lady Jane decided to go for the walk. Jane got up quietly. Jane thought that Edward was asleep. Jane got dressed. Jane went to the garden. Edward followed Lady Jane. Jane saw Edward. Lord Edward had a long dagger. Edward waved the dagger wildly. Lord Edward stabbed Jane screaming. The knife sank deep. Jane struggled weakly. Jane hit Edward. Lord Edward slashed Jane again. Edward said that Lady Jane betrayed Lord Edward. Jane dying covered with the blood. Lord Edward hid the knife. Edward returned to the bedroom. Lord Edward washed off the blood. [klein1973automatic, p. 109]

 Though briefly identified by Ed Kahn [kahn1974automatic, p. 5] and then later by Margaret Boden [boden1977artificial, p. 300] and Mark Lee [lee1994model, p. 19], Marie-Laure Ryan has probably provided the most extensive articulation of this issue of premature revelation in the generated “mysteries”. Here is an excerpt:

> Since the murder is narrated before it is solved the stories lack suspense, the essential ingredient of the genre they are supposed to represent. The program produces the historical sequence of events underlying a detective story, but not a well-formed textual rendition. [ryan1987heuristics, p. 515]

 Whenever there are interesting simulated events but a misshapen narrative account of them, poor curation is to blame. Thus, Klein's system's deficiencies are not rooted in simulation—though the modeling gap of a lack of detection is another critical blunder, as discussed in Section 4.1.5—so much as in curation. The stuff of a good mystery story is produced by the simulation, but the system does not have the means to tell a good mystery, because it does not know how to arrange a subset of the generated material into the mystery emplotment. [Note: Here, I take the liberty of using Hayden White's term ‘emplotment', which was discussed at length in Section 3.1.1.]

#### Curation and Tellability

Like all narrative devices, the mystery emplotment's reliance on suspense connects back to the work surrounding *tellability* and *story interestingness* (particularly *cognitive interest*), which I discussed in Section 4.1.1. David Elson made this connection in his dissertation:

> the creation of mystery is a device to increase the “tellability” of the story by prompting the receiver to have its own goal—to achieve causal closure in its cognitive model of the story's meaning. As receivers, we are given pieces of a large puzzle, one at a time and in a non-random order, and we must find the most likely (or most satisfying) assemblage of all the facts into a coherent whole. [elson2012modeling, p. 305] [Note: A quotation of China Mieville that I recently encountered in Aaron Reed's dissertation [reed2017changeful, p. 98] provides another evocative articulation: “These are novels of potentiality. Quantum narratives. Their power isn't in their final acts, but in the profusion of superpositions before them, the could-bes, what-ifs and never-knows. Until that final chapter, each of those is as real and true as all the others, jostling realities all dreamed up by the crime, none trapped in vulgar facticity” [mieville2009crime, n.p.].]

 Thus, while the murder event is itself inherently tellable—following the discussion in Section 4.1.1, murder is an ‘absolute interest' [schank1979interestingness], ‘human dramatic situation' [wilensky1982points,wilensky1983story], ‘generically important element' [freebody1986serial], and an event of ‘situational' [krapp1989interest,hidi1990interest,hidi1992situational] and ‘instinctive interest' [behrooz2018cognitive]—but the other story events are tellable primarily for being causally related to the murder.

Thus, as the crux of the murder mystery, the murder itself is what William Labov would call the *most reportable event*, as discussed in Section 4.1.1. In describing the act of “narrative pre-construction” that is carried out by storytellers in everyday conversational narrative, Labov informally provides an algorithm for what I call *story sifting* (note that I have already quoted this above):

> Pre-construction begins with this most reportable event and proceeds backwards in time to locate events that are linked causally each to the following one, a recursive process that ends with the location of the unreportable event—one that is not reportable in itself and needs no explanation. [labov2006narrative, p. 37]

#### How to Sift Story Sequences

Labov's procedure could be carried out automatically in Klein's system by starting with the murder event and then chaining backward along the causal sequence of events that precipitated the killing. Per the discussion in Section 4.1.6, such a causal sequence will specifically be constituted according to *contingency* relations. As I explained there, narratologist Mark Alan Powell identifies contingency as obtaining between two events when one clearly and explicitly causes the other.

In the case of the story whose excerpt I quoted above, backchaining from the murder (most reportable event) across contingency links would lead back to actions including Lord Edward becoming aware of Lady Jane's affair, Lady Jane having the affair, Lady Jane's flirtations that preceded the affair, and the events that lead to the social practice in which those flirtations occurred (e.g., tennis game), and so forth, finally terminating in an action at the beginning of the simulation. In this way, the system would be able to *sift out* (from the larger accumulation of all simulated events) the emergent contingency structure that constitutes the story's *plot skeleton*. In identifying such sequences, a system may automatically distinguish between what Roland Barthes calls ‘cardinal functions' and ‘catalyzers' (these are the conventional English translations):

> [narrative] units are not all of the same “importance”: some constitute real hinge points of the narrative [while] others merely “fill in” the narrative space separating the hinge functions. Let us call the former *cardinal functions* (or *nuclei*) and the latter, having regard to their complementary nature, *catalyzers*. [barthes1975introduction, pp. 247–248]

 In translating Barthes, Seymour Chatman instead uses the English terms *kernel* and *satellite*: [Note: I prefer these terms, so I will use them instead of ‘cardinal function' and ‘catalyzer', the latter of which Chatman argues is a mistranslation [chatman1980story, p. 54].]

> Kernels cannot be deleted without destroying the narrative logic. In the classical narrative text, proper interpretation of events at any given point is a function of the ability to follow these ongoing selections, to see later kernels as consequences of earlier. A minor plot event—a satellite—is not crucial in this sense. It can be deleted without disturbing the logic of the plot, though its omission will, of course, impoverish the narrative aesthetically. [chatman1980story, pp. 53–54]

In the case of emergent narrative, the satellites of a given kernel sequence could be any action that occurs in the background of the simulation. Because simulations model the storyworld with much higher level of detail than print fiction or other conventional narrative forms—this was the crux of my argument in Section 3.1.1 that emergent narrative works like nonfiction—events may occur that have essentially nothing to do with a given kernel sequence. In this case, such an event should probably not even be considered a satellite of that sequence, but rather an occurrence totally outside its orbit. Moreover, in the kind of massive emergent narrative that I am advocating, where many characters live out entire lives in the world, there is a major possibility of multiple kernel sequences emerging, potentially with some intriguing emergent entangling. Indeed, this prospect of many interrelated storylines emerging is the strongest selling point of the approach. Thus, while Barthes argues that there is a distinct kernel sequence at the core of a given narrative world, as it is reported in a work of fiction, in the case of simulation we might consider kernel sequences to emerge whenever a tellable event emerges. That is, when a tellable event emerges, we may trace is causal lineage (in terms of contingency relations) to excavate a series of kernels that constitute an emergent storyline that is indexed by the tellable event, which in Labov's terminology constitutes the most reportable event in that storyline. [Note: Of course, if a more tellable event is encountered through the excavation of the causal structure, then the story may be redefined as being about *that* event.] While Labov's algorithm implies that a story will always culminate in its most reportable event (since the process of excavation is one of backward chaining), this is of course not a real constraint—indeed, by forward chaining across contingency links, a system may also recognize the ramifications of the most reportable event.

Note that the excavated kernel sequence (a plot skeleton) may actually take the form of a graph, rather than a linear sequence or a tree. This is because the kernel nodes (actions or statives) may have multiple parents (actions or statives that caused them) and multiple children (actions or statives they caused). [Note: As I discussed at length in Section 4.1.6, in the real world and in a simulation, the true causal structure may actually take the form of a densely interconnected lattice: since all actions change the world and the state of the world determines whether any action may be taken at any time, each occurring action may be viewed as being caused by each and every earlier action. The creators of *Prom Week*, for example, implicitly describe such a causal structure in their articulation of *social physics*: “In the delicate system of social physics, the smallest social change reverberates and impacts the entire system” [mccoy2011prom, p. 319]. However, humans seem to adopt simplified models of causality that resemble graphs, rather than densely interconnected lattices, and certainly this is how causality works in stories, as narratologists like Barthes and Chatman and Powell have shown. Thus, to do story sifting properly, a system must be able to reason about causality such that excavated causal structures are topologically on the order of the causal structures that humans tend to work with when constructing and understanding stories. Following Powell's distinction, this means recognizing contingency links, rather than links that capture the causality relationships of possibility and probability.] Marie-Laure Ryan, who has also emphasized this distinction between a plot skeleton and the larger accumulation of narrative material in which such skeletons are embedded, references this graphical structure:

> Within the narrative universe, we can distinguish between a raw historical sequence, consisting of all the physical states and events in their chronological order, and a rationalized sequence, in which information is integrated into a graph according to the rules of concatenation. In this second sequence, each action is linked to a motivating intent, and each intent can be traced back through a causal chain to a physical state or event. [ryan1986embedded, p. 322]

 In Ryan's formalism, which is introduced in that paper, nodes in the graph may be linked according to relationships beyond contingency. Indeed, in the case of automatic curation, a system could use whatever semantics it wishes to label the edges according to anything that may be of use to the presentation module that will be generating an actual narrative account of the graph. Of course, it would also be reasonable to include satellites (using edges that express as much) and any other material that could be of use to the presentation module.

For such purposes, one might harness the expressive power of the *story intention graph*, David Elson's graph-based story schema that I mentioned in Section 4.1.1. In recent work by my colleagues at UC Santa Cruz, the story intention graph has been used as a representation scheme that enables generative story *re*telling [rishes2013generating,lukin2015narrative,lukin2015generating,lukin2017generating] and richer case-based reasoning for *Minstrel*-style story invention [harmon2015imaginative,harmon2015revisiting,harmon2017narrative]. [Note: The formalism's utility for carrying out automatic reasoning about *story analogy* was established by Elson himself [elson2012detecting]. Inspired by the work of Keith Holyoak and Paul Thagard [holyoak1989analogical], Elson's approach operationalizes analogy as subgraph isomorphism.] Thus, one specific prospect of automatic curation is to sift through the accumulation of material produced by a simulation to excavate kernel sequences that may be used to automatically build story intention graphs, which could then drive the generation of narrative accounts that are expressed in natural language. This was the ultimate goal of my collaborative work on generating *Prom Week* narrative postmortems [antoun2015generating], though we never got past a proof-of-concept stage in which we resorted to the manual construction of story intention graphs from playtrace data. [Note: This was due in part to the project leads, Chris and Matt Antoun, departing UC Santa Cruz after graduating with their masters degrees.]

#### Fixing Klein's System

Let us now return to Klein's “mystery” generator. While it has no implemented means with which to excavate kernel sequences, there is a technical solution that is made trivial by the elegant causal feedback loops that drive the system's emergent narrative. As I explained in Section 4.1.6, the unfolding of murder stories is critically enabled by a series of *story registers* that track important information about what has happened so far. For example, when two characters flirt, they are added to a register that enables a future tryst between them. Likewise, when a series of emergent actions results in a character adopting a motive to kill someone else, a register tracks this, which unlocks the possibility of that character murdering the prospective victim later on.

While these registers are used in a *forward* direction to structure the emergent possibility space, they could also be used in a *backward* direction to enable story sifting. To do this, the system would essentially chain backward across the registers, starting from the murder: retrieve from the motive register the entry that groups the murderer, victim, and motive, then proceed to the tryst register and do a similar lookup before likewise chaining backward to the flirtation register, and so forth. In some cases, it might not be possible to chain from one register to the causally preceding one, but this can easily be rectified by including the linking actions in the entries that are added to the registers—that is, as the simulation is unfolding, add into these registers the information that will be needed to chain backward during story sifting after the simulation concludes. I call this task *causal bookkeeping*. In Klein's system, this would mean adding into each entry in the murder registry a pointer to the flirtation-register entry that enabled the former, and so forth for all such contingency links.

Even if Klein's system could utilize causal bookkeeping to do story sifting, simply recounting an excavated kernel sequences in chronological order may not produce a good telling. For one, presentation is incredibly important, as I discuss in the next section, which means that bad prose will ruin otherwise good stories. [Note: Will good prose save otherwise bad stories? I want to explore this notion in a study. My idea: have great writers write short stories that recount the most boring *Tale-Spin* outputs, and see how readers rate them relative to the originals. It would also be interesting to explore the inversion of this effect by rendering a beloved story in the style of a *Tale-Spin* output and seeing what people make of it. In a 1982 paper, Meehan himself acknowledged this dimension (following Natalie Dehn's emphasis on the author [dehn1981story]): “In artificial intelligence terms, what we have here is a domain problem. We are trying to reason about a form of writing, without having much expertise in the art of writing, and I expect no more success than I would from someone with no medical knowledge who tried to build a diagnosis system. The [story generation] literature reveals little or no contact with the practitioners, namely writers” [meehan1982stories, p. 459].] More immediately, though, the *arrangement* of the selected material in the actual telling is also critical. Here, we encounter the classic distinction between *fabula*, the true temporal ordering of selected events, and *syuzhet*, the ordering of those events in the telling. [Note: In the case of historiography and curationist emergent narrative, we might also identify a third layer, which is the ordering of all events that occurred in the world, regardless of whether they were selected for telling. In the case of conventional fiction, this distinction is coextensive with the fabula, since the only events that can be known to have occurred in the fictional world are the ones that are reported in the work. This distinction is another argument for why emergent narrative works like nonfiction.] As established above, the fatal flaw of Klein's generated murder mysteries is that they are not mysteries at all, because the identity of the murderer is revealed as the crime is transpiring. In a proper mystery story, there is divergence between the fabula and syuzhet such that the mystery is resolved later in the latter. Such divergence requires a decoupling of fabula and syuzhet, which in the case of emergent narrative requires a decoupling of simulation and telling. In Klein's system, however, the story is literally told *as it is happening*: the rules for natural language generation define which events should be recounted, and those events are recounted *as they occur*. Thus, the telling is actually a kind of simulation trace, but it does not recount every single event that occurs, so it is not monolithic like the extreme case discussed in the last section.

There are two solutions to the problem of introducing suspense into Klein's generated stories; one of these is specific to his project and the other is a general solution. The first solution, by which Klein's system could be “fixed” in less than an hour, is to simply change the rules for generating the natural language that expresses the actions that make up the recounted murder scenes. Recall that prior to a murder occurring, it is likely that several characters will have adopted motives for killing other characters. Thus, when a given character is murdered, it is likely that multiple other characters will have adopted a motive for killing that person. As Marie-Laure Ryan notes, “a good mystery story should implant the false clues of possible, but unrealized explanations” [ryan1987heuristics, p. 515], and this is exactly what the earlier scenes do.

Since there is no exclusive connection by which the murderer can be identified given a victim (which would be the case if only that character had a motive for killing the victim), suspense could be generated simply by rendering the murder scene without revealing the murderer's identity. This could be achieved by changing the rules for natural language generation (for the actions associated with the murder scene) such that identifying noun phrases are replaced with constructions like ‘the killer' and pronominal expressions with non-gendered variants. Because the suspense depends on multiple characters having a motive to kill the victim, the simulation might also need to hold off on the murder scene until such circumstances hold for a potential victim. Similarly, Marie-Laure Ryan proposes a reformulation to Klein's system such that narration only begins after the murder scene, whereby the stories would be all about the detection phase [ryan1987heuristics, p. 515]. [Note: Similarly, Ed Kahn stated that a better telling could be achieved if a “global editing program [...] filtered out excessive randomness and gave more coherence to the story as a whole.” [kahn1974automatic, p. 5]. Below, I discuss how Klein himself actually intended for such curation.] This would take considerable work to implement well, however, due to the system's lack of rich detection modeling, as discussed in Section 4.1.5.

A more general solution to the introduction of suspense characterizes a broad approach to the task of curating simulated storyworlds: decouple the simulation and the curation procedure, such that curation occurs *after* the simulation has concluded. In the case of Klein's system, curation would begin once the simulation has concluded. This would entail the development of a second system that identifies the murder action as a most reportable event, and then chains backward and forward to excavate the larger kernel sequence containing the lead-up to the crime and also its ramifications. This would result in a lot of extraneous material being excluded, which would generally help the prose, but it would also make sense to include satellites (to use Chatman's term) pertaining to red herrings such as non-murderers adopting motives to kill the victim. Such excavation of material beyond the kernel sequence indexed by the most reportable event is the real challenge of curation, since it requires an author to inscribe knowledge about (good) storytelling beyond the relatively simple matter of assembling a sequence of temporally and causally related events.

#### Klein Cared About Curation

Before moving on, I would like to clarify that Sheldon Klein was actually quite concerned with the idea of curation, which he referred to as “narrative style control” [klein1971program, p. 20]. In line with my proposal for automatic curation, Klein's *narrative style control monitor* was to reason not just about the style of prose, but also about the selection of material to recount: “The problem is what to describe”, he aptly observed [klein1971program, p. 20]. In his first note on the project, Klein proposes that the narrative accounts could be limited to one character's experience:

> Design of the system includes the ability to follow just one character and what impinges upon him through the course of the plot simulation, or all characters simultaneously, or serially. [klein1969automatic, p. 418]

 This selection technique, coupled with Klein's plans to generate first-person prose in targeted character voices [klein1973automatic, pp. 28–29], could have enable automatic *focalization*, which was not explored in story generation until several decades later [callaway2002narrative,lonneker2005narratological,cheong2006computational,montfort2007generating,bae2011automated].

Moreover, Klein writes lucidly about how it is necessary in any case to select a subset of simulated material for recounting:

> One might use the total semantic network to generate a total description of itself each and every time frame at a horrendous cost in time and redundancy. The first version of this program was set to issue descriptions of just the changes in the semantic network that occurred during the previous time frame. [...] To control the subject matter, it is necessary to have a powerful device for selecting particular subpaths through the semantic network to serve as inputs to the narrative [i.e., narration] generation component. Accordingly, a program is under construction for finding complex paths through the network that are a function of a variety of logical conditions. Typical requests to the program might be: “Find a path between node A and node B that passes through relation $R_1$, but not through $R_2$.” [klein1971program, pp. 20–21] [Note: A similar proposal appears in his other paper on the system [klein1973automatic, p. 28].]

 Indeed, such requests could be used to isolate arbitrary subgraphs of the larger semantic network that Klein uses to represent the storyworld and its accumulated history. In this way, an author could inscribe patterns about what makes a good story. This expresses what I view as a strength of Klein's perspective, as opposed to the more or less contemporaneous *Tale-Spin*: Klein views story generation as a kind of latent authorship, while to Meehan it is an experimental method for doing cognitivist research that is essentially divorced from aesthetic concerns. Klein even mentions, humorously, that “the task of writing rules for generating a novel will probably involve as much work as the old-fashioned method” [klein1971program, p. 31].

#### Coining ‘Curationism'

Thus, we find that the idea of curation is likely as old as the idea of emergent narrative: as I have shown, Klein described curation at some length, and his critics pinpointed poor curation as the fundamental limitation of his system. However, while these ideas have been in place for a long time, there has yet to be an extensive articulation of the act of curation as being *the central matter of importance in emergent narrative*, beyond even simulation. Such articulation is the fundamental contribution of this dissertation, and it is why I use the neologism *curationism*: in my framework, curation is emphasized above all else.

### Poor Presentation

As the previous sections show, practitioners of emergent narrative have often been more concerned with the generation of narrative raw material than with any subsequent curation of that material. One particularly salient form of curation error pertains to the quality of presentation of a narrative artifact; in the history of story generation, this has typically manifested as bad prose. Unfortunately, is not hard to find examples of this kind of deficiency.

#### *Tale-Spin 's Stilted Prose*
Though it is not alone in this category, *Tale-Spin* is a prime example of poor presentation—in its case, bad prose—doing considerable harm to a project that is in many other ways quite impressive. This failing of the system is already well established (e.g., [boden2015foreword]), but let us briefly take a look at the opening excerpt of an example generated output:

> Once upon a time George Ant lived near a patch of ground. There was a nest in an ash tree. Wilma Bird lived in the nest. There was some water in a river. Wilma knew that the water was in the river. George knew that the water was in the river. One day Wilma was very thirsty. Wilma wanted to get near some water. Wilma flew from her nest across a meadow through a valley to the river. Wilma drank the water. Wilma was not thirsty. [meehan1976metanovel, p. 227–228]

This prose is clearly of poor quality, and for our purposes here it is not worth getting into the specifics, since this is not a thesis on natural language generation. Rather than discussing what makes this prose bad, I think it is more interesting to consider why this prose is bad. As Meehan makes clear in his descriptions of *Tale-Spin*, text generation was an ancillary concern in this project, which is primarily an exploration of character planning as a model for human cognition. Originally, Meehan had hoped to use a natural language generation system called *Babel* that was developed by another Roger Schank student, Neil Goldman, as part of his own dissertation work [goldman1974computer]. *Babel* unfortunately took up too much memory and was also missing a few features that Meehan required, so he decided to instead quickly hack together his own system, which he (aptly) called *Mumble*. [Note: Meehan's *Mumble* is not to be confused with David D. McDonald's *Mumble*, which is one of the major early systems in natural language generation [mcdonald1980natural,mcdonald1981mumble,mcdonald1983description,mcdonald1985description], and one that is still in use today [mcdonald2018mumble].] Here is how Meehan describes *Mumble* in his dissertation:

> It's a quick ‘n' dirty program, written in a day, and it [sic] many of its parts do not correspond to the way humans speak, but it produces adequate, if somewhat verbose, sentences. Best of all, it's one tenth the size of BABEL. [meehan1976metanovel, p. 200]

 Given the intellectual and technical contexts of its development, it is not surprising that *Tale-Spin* features bad prose—presentation was not a focus for Meehan.

#### Klein and the Cutting Edge

Another early project, however, features bad prose for a different reason. As noted above, Sheldon Klein was a pioneer of natural language generation who pushed its cutting edge throughout the 1960s and early 1970s [simmons1962toward,klein1963syntactic,klein1965control]. Thus, while the prose produced by his murder-mystery generator is of dubious aesthetic quality, [Note: The scholar Bernhard Lindemann characterized it as evoking “something like ‘The New York Telephone Directory Retold and Edited by Gertrude Stein'{”} [lindemann1983text, p. 5]—that, however, sounds like an intriguing work to me!] it actually represents the state of the art in natural language generation for its time. For instance, it was only in 1961 that Victor Yngve astounded the nascent world of artificial intelligence by generating random sentences that were syntactically well formed, such as this one:

> When the water under trains has oiled wheels, four polished, shiny, little, oiled and little fire-boxs [sic] and its black, polished, proud, big, black and oiled fire-boxs, Engineer Small keeps the four wheels and fire-boxs black, big, little, oiled and heated. [yngve1961random, pp. 77–78]

 By the early 1970s, when Klein published outputs from his system, the state of the art had not changed by much, and Klein was himself a leading figure in computational linguistics.

Thus, we might look a bit more deeply (and with more sympathy) at his generated prose, such as this excerpt here:

> James knew that Hume screwed Marion. James hated Dr. Bartholomew Hume. James wanted a revenge [sic]. James decided to kill Dr. Hume. [klein1973automatic, p. 108]

 Note how the character Dr. Bartholomew Hume is referred to using three different noun phrases—this reflects effort on behalf of the system (and on behalf of Klein and his students as its authors) to feature lexical variation in the generated prose. In another excerpt from the same story, we find similar variation (‘the policemen' vis-à-vis ‘the cops') and also the impressive sentence “Florence also looked”, which is a paraphrase of ‘Florence looked for clues in the library' enabled by anaphora:

> The policeman arrive. The cops were idiotic. A detective examined the corpse. The cops looked for clues in the library. Florence also looked. [klein1973automatic, p. 108]

In fact, Klein designed his system to provide authorial control along several dimensions of curation (particularly with regard to presentation), which he identifies as including “lexical frequency, syntactic structure frequency, narrative subject matter, descriptive complexity, internal paragraph structure, paragraph grouping” [klein1971program, p. 20]. [Note: Section 5.1 of Klein's second paper on the system further emphasizes these architectural affordances [klein1973automatic, pp. 27–29].] Moreover, by this time Klein had already demonstrated the feasibility of exerting authorial control along these dimensions in what is perhaps the earliest work in natural language generation that emphasized authorial control of style [klein1965automatic,klein1965control]. The problem, however, was that the specification of such control mechanisms would have been particularly burdensome in this era of computer programming. Thus, Klein's dubious prose does not reflect a lack of interest in natural language generation—he was a leader in the field and he situated the project as a contribution to that area—but rather they reflect the state of the art at that time with regard to the collected knowledge of its practitioners and moreover the hardware limitations of the era. As Klein explained in a later paper, the speed at which his system generated its prose was a technical achievement in that time period:

> Achievements with the generative portion of the system include a text grammar model that generates 2100 word murder mystery stories in less than 19 seconds each, complete with calculation of the plot and specification of the deep structure as well as the surface syntax. The speed of this generation is 100 to 1000 times faster than other existing programs using transformational grammars. (The algorithm for the semantics-to-surface structure generative component is such that processing time increases only linearly as a function of sentence length and syntactic complexity.) [klein1975meta, p. 84]

#### Hardware Limitations and Emergent Narrative

While hardware limitations are no longer a major concern in natural language generation, it is important to consider them when discussing historical projects such as Klein's and Meehan's. More broadly, we still face limitations today in terms of working memory and computational efficiency, and these do constrain work in emergent narrative. A good example of this is the use of *level-of-detail modulation* in underlying simulations [doull2008death,mark2009aiide,ohalloran2016multi]. [Note: Klein incidentally appears to have proposed level-of-detail modulation in a discussion of prospects for future work: “Other features include variable time scales for different rules and, accordingly, variable subdivision of the major time frame into subtime frames” [klein1971program, p. 32]. Curiously, Klein notes this about such modulation: “Interestingly, this feature creates problems that tempt one to assume Einstein's view of time and space; discussion of that topic is reserved for another paper” [klein1971program, p. 32]. Unfortunately, to my knowledge such a paper never appeared.] In *Dwarf Fortress*, for instance, areas of gameworld are modeled more abstractly and simulated more coarsely depending on how far away they are from the player. [Note: Personal communication with Tarn Adams, June 15, 2018.] Another kind of approach in this vein pertains to *supporting character realism*: modeling supporting characters (typically in a real-time graphical world) in a computationally efficient way that does not break immersion [jan2012virtual,skorupski2012three,fallon2013believable,eccles2017procedurally,plch2018belieavable]. As Part II of this dissertation will show, I tend to eschew this technique in my own work, since I see it as a form of simulation tampering that inhibits the pleasures of emergent narrative outlined in Chapter 3, but this is at the cost of computational efficiency.

To return to Klein: while his system's generated prose is clearly of low aesthetic quality, this is not due to apathy or even to poor execution. Klein cared about presentation and he set out to produce prose of high quality, but it was simply not feasible to do so at this time. He could have decided to abandon the project early to work on something more near-term, but instead he forged onward to materialize the first widely reported effort in story generation. There will always be technical and epistemic limitations at play in the development of works of computational media. Practitioners simply have to work within these confines, but hopefully in doing so the confines will loosen. In considering historical works, especially computational ones, it is important to consider the particular limitations that would have constrained the development of the work in its time.

#### Emphasizing Emphasis

When it comes to presentation, I think it is probably most rewarding to consider how much a creator appears to have cared about it, and how much effort they exerted in attempting to produce artifacts that are presented well. Jim Meehan did not care about presentation, and the prose generated by *Mumble* reflects this. Sheldon Klein, on the other hand, did care about presentation, but unfortunately only a deep examination of the prose (informed by historical context) reveals this emphasis. [Note: Noah Wardrip-Fruin has called the practice of studying works of computational media with attention to their underlying processes and historical contexts *digital media archaeology* [wardrip2011digital]. Note: I prefer the term *computational media archaeology*, since computational processes are not inherently relevant to the understanding of all digital media (e.g., a digital image or a digital audio file).] Some projects in story generation have been concerned entirely with presentation, or more specifically, with generated story prose. To my knowledge, the earliest example of this kind of effort is Paul Juell's forgotten 1981 dissertation, in which he seeks to improve the quality of prose for stories generated by Klein's system, *Tale-Spin*, and *Telltale* [juell1981improvements]. [Note: Like Ed Kahn's 1973 PhD thesis [kahn1973foundations], Juell's Ohio State dissertation is also among the earliest in the field of story generation and it has likewise been forgotten to time. Interestingly, I recently encountered a 1993 post to a newsgroup in which Mark Kantrowitz, who worked on the Oz Project [kantrowitz1990glinda,kantrowitz1992integrated], references Juell's dissertation as one of the few prior efforts in stylistic natural language generation (a tradition that he connects back to Sheldon Klein's 1965 paper “Control of Style with a Generative Grammar” [klein1965control]) [kantrowitz1993style]. Indeed, Kantrowitz's massive bibliography of natural language generation [kantrowitz1993bibliography] is one of two documents to cite Juell's dissertation, the other being a more famous Ohio State thesis: Brenda Laurel's [laurel1986toward, p. 144].] For instance, here is an example of *Tale-Spin*'s default *Mumble*-based output:

> Once upon a time George ant lived near a patch of ground. There was a nest in an ash tree. Wilma bird lived in the nest. There was some water in a river. Wilma knew that the water was in the river. George knew that the water was in the river. [meehan1976metanovel, p. 237]

 For comparison, this is Juell's system's improved prose for the same passage:

> Once upon a time George Ant lived near a patch of ground. There was a nest in an ash tree and some water in a river. He knew that the water was in the river. Wilma Bird knew that the water was in the river. She lived in the nest. [juell1981improvements, pp. 176]

 Later efforts to improve the prose of story generators (or target particular pragmatic effects) include Ed Hovy's *Pauline* [hovy1987generating,hovy1987thesis], Mark Kantrowitz's *Glinda* [kantrowitz1990glinda,kantrowitz1992integrated], Mark Lee's unnamed 1994 system [lee1994model], Clark Elliot's *story-morphs* project [elliott1998story], Charles Callaway and James Lester's *StoryBook* [callaway2002narrative], Nick Montfort's *Curveship* [montfort2007generating,montfort2011curveship] and *Slant* systems [montfort2013slant,montfort2014expressing], and Stephanie Lukin's *Fabula Tales* [lukin2015generating,lukin2015narrative,lukin2017generating].

#### Presentation Modalities Beyond Prose

Finally, I would like to note that narrational prose is not the only presentation modality that is possible for emergent narrative. Later on, I will discuss a number of potential alternatives, ranging from physical installations to enacted performance to conversational speech to videogame quests to comics and more. Interestingly, some of these modalities appear to be easier to work in, in terms of the amount of limitations placed on the aesthetic quality of the presented artifacts. In any case, however, constraints are placed on each modality, and a lack of emphasis on presentation will doom works of emergent narrative regardless of the particular medium that is targeted for a curated narrative artifact.

#### Closing

To summarize the remarks of this section, poor presentation in earlier projects seems to have often resulted, naturally, from a lack of emphasis on the presentation aspects of a project. That being said, even when presentation is stressed, technical and epistemic deficiencies may limit the potential aesthetic quality of a presented artifact, and it is especially important to keep this in mind when considering historical works in the area of computational media.

### Failure to Mount

One aspect of presentation is important enough to merit its own section, especially considering the degree to which it has contributed to the poor showing of emergent narrative to date. In this dissertation, I assume that projects in emergent narrative are more broadly concerned with media, which means that the practice is really about developing media works. [Note: As noted in the previous section, practitioners like Jim Meehan have not always been concerned with media. In those cases, the works do not even make it far enough to fail in the particular way described in this section, and thus they are not worth considering here.] A major shortcoming of the work in this area, particularly projects that have come out of academia, is that its practitioners do not seem to emphasize the construction of actual media artifacts. That is, even projects in which curated narrative artifacts may be generated fail to then *mount* these artifacts as (or in) full-fledged media works.

#### Stop Mounting in Academic Papers

As a fundamental example, consider how generated artifacts in story generation, even ones that are the result of some degree of curation, have typically been presented: prose that is included in a larger academic paper. This is a not an authentic (or at least compelling) media genre. [Note: And I say this as one of the world's major connoisseurs of generated stories printed in academic papers.] While I concede that there is a certain aesthetic appeal to generated stories that are printed in all caps on fanfold printout—this is how *Saga II*'s raw outputs appear, as Figure 4.5 shows—the attraction is probably rooted in a kind of historical fetishism. This is not really a real media genre, is it? [Note: A potential counterexample here is Alison Knowles and James Tenney's *A House of Dust*, which is an early project in computer poetry that resulted in the publication of a chapbook that was distributed as fanfold printouts contained in plastic bags [knowles1970house]. This kind of presentation is evocative by virtue of its emphasis on the materiality of the computer poetry of its era. That is, fanfold printout is the final form of the curated product as the result of conscious deliberation—a kind of aesthetic doubling down—not simply because fanfold printout is how computer outputs were often experienced at that time. Moreover, the artifacts were distributed, which reflects an emphasis on culminating in a media artifact.] I contend that projects whose curated narrative artifacts are presented in this way fail to complete the process of media creation.

[Figure omitted]

#### Klein's Research as Media Practice

Sheldon Klein realized this. In his first technical report on the murder-mystery generator, he wrote the following about his aims for the project:

> It is a rather long distance from [the] promise of the title to the fragmentary output just presented. The first author of this paper is an outspoken logical positivist, and accordingly, foresees, as the inevitable test of the system and the semantic model it embodies, the publication of a novel produced by the program, with the publisher remaining ignorant of the non-human origin of the material. [klein1971program, p. 30]

 Klein articulates this goal using the language of scientific investigation, but I believe that this passage also suggests the importance of mounting generated outputs in full-fledged media works—in this case, a novel. After all, he calls his system an “automatic novel writer”, not a ‘story generator'. [Note: Klein uses this phrase as early as 1969, in an interesting note on the project that reports a start date of September 1967 [klein1969automatic, p. 418]. In terms of the history of story generation, this is is considerably earlier than 1971–73, which is the period with which scholars who acknowledge the project have associated it. Still, it is not the earliest work in computer story generation, as I have already noted above: *Saga II* (1960) and a folktale generator by Joseph E. Grimes (circa 1963) preceded it [ryan2017grimes].] To be fair, the latter phrase was not yet in use, but in any event Klein makes it clear that his project targets the specific media form of the novel.

Though his novel was never produced, it appears that Klein came close, as he wrote in the April 1974 SIGART Newsletter, “Results of our efforts should be available by the end of Summer 1974 in the form of a first draft of a book entitled, appropriately, The Machine that Ate the World” [kahn1974automatic, p. 4]. [Note: Note that I call dibs on using this title at some point, potentially for my planned book on Sheldon Klein and his wild career.] While Klein's book never appeared, the system's outputs did appear in Jasia Reichardt's *Robots: Fact, Fiction, and Prediction* [reichardt1978robots, pp. 70–72] and in the French literary magazine *Action poetique* [balpe1984action, p. 33]. [Note: A decade earlier, the linguist Joe Grimes had planned to publish a book of computer-generated stories, which he would have called *Grimes' Fairy Tales*, but it too never appeared [ryan2017grimes]. I have encountered evidence that *Saga II* cocreator Douglas T. Ross had several of the system's outputs compiled into a one-off bound volume, but I do not know if this artifact survives. Over the remainder of the century, at least three published novels would be marketed as having been written by computer programs: the obscure *Bagabone, Hem'I Die Now* (1980) by *Melpomene* [melpomene1980bagabone], the famous *The Policeman's Beard is Half Constructed* (1984) by *Racter* [racter1984policeman], and the remarkable *Just This Once* (1993) by *Hal* (“as told to Scott French”) [french1993just]. While the former two works have been exposed as hoaxes [schultz1980publisher,barger1993policeman], French's admitted collaboration might have authentically integrated significant computer contributions. In the book's introduction, French claims to have spent $$50,000$ on professional-grade *expert systems* technology [french1993just, pp. vii–viii], which he used to author thousands of production rules that operationalized the style of writer Jacqueline Susann [boudreau1993romance]. The book has been the subject of academic legal discussion [vigderson1994hamlet,clifford1996intellectual,wu1997video,barfield2006intellectual], and even a lawsuit: Susann's estate contended that the system plagiarized her and sued French for copyright infringement, but the parties settled out of court [lohr1993potboiler]. Just last year, an indubitable volume of computer-generated stories was published by Counterpath Press: Rafael Pérez y Pérez's *Mexica: 20 Years—20 Stories [20 anos—20 historias]* (2017) [perez2017mexica], which compiles twenty tales produced by his *Mexica* system [perez1999mexica]. Finally, many volumes of computer poetry have been published over the decades.] In that same SIGART Newsletter in which his forthcoming book is announced, Klein also alludes to another evocative project: “There is even work on a musical composition generator based on a narrative discourse generation model” [kahn1974automatic, p. 4]. This thread, it appears, would evolve into an effort to generate the multimedia video opera *Revolt in Flatland*. Here is a description of the project that appeared in *Byte Magazine*:

> More recently, Klein says, he and his students have invented a program to write operas. Like the earlier two programs, the opera-writer is divided into two parts. A simulator generates the plot in a special symbolic language devised by the group, then translators turn the symbolic plot into words, music, and moving pictures of the action. [...] To facilitate display on a video terminal, the opera is based on Edwin Abbott's nineteenth-century fantasy *Flatland* about a two-dimensional world [mckean1982computers, p. 51]

 In a brief 1982 note, Klein notes this about the project:

> We use the simulation model as a generator of operas, complete with textual, pictorial and musical output, all derived from a common semantic source. The output is in the medium of a videotape recording, and is generally entitled, *Revolt in Flatland*. [klein1982propositional, p. 150] [Note: I do not know of any earlier examples of algorithmic scoring in a multimedia work. Later, the Oz Project spinoff company Zoesis produced *The Penguin Who Wouldn't Swim* (1999) [zoesis1999penguin], an interactive drama (in the style of a Pixar short) featuring an adaptive score implemented by composer Michael Gordon Shapiro [shapiropenguin]. The project was informally presented at the 1999 AAAI Fall Symposium on Narrative Intelligence, as Noah Wardrip-Fruin notes [wardrip2009expressive, p. 329], and it was also briefly mentioned in a 2004 note by Bryan Loyall [loyal2004response]. In a recent conversation, Michael Mateas called it the most impressive piece to come out of the Oz/Zoesis effort.]

I appreciate how this description is clear about the specific target medium, opera, as opposed to the nebulous ‘story,' which enables the cop-out of merely printing generated prose in an academic paper. It would be ridiculous to claim that an opera has been produced by its reporting in a document, but doing this for a ‘story' seems to command less scrutiny. Moreover, Klein does not just state a target medium, but also a target media format (‘a videotape recording'). When it comes to this project, Klein appears to have achieved his goal: a *Revolt in Flatland* videotape was shown at an academic conference [mckean1982computers], and ultimately the work was included in the Central Opera Service directory of operas that were composed in the 1980s [opera1990directory, p. 72].

#### Narrative Artifacts Are Not Enough

As others have articulated elsewhere, computational systems that are intended to support media experiences cannot be truly appraised except through actual implemented experiences of the kind that they are meant to support [mateas2001expressive,stern2008embracing,horswill2013science,koenitz2014five,wardrip2014envisioning]. In the case of emergent narrative, it is possible for a project to feature a strong simulation and sound curation, resulting in compelling narrative artifacts that still disappoint, due to not being mounted in an actual media experience. This can be caused by its practitioners not having a clear sense of the medium that they are targeting (beyond ‘story'), and of course it is the unfortunate case of all projects that do not quite reach the finish line of culminating in actual media.

Before concluding this section, I would like to briefly note that the narrative artifacts created in projects in emergent narrative do not have to themselves be mounted as standalone projects. Indeed, there is the intriguing prospect of mounting the artifact as a component in a larger media experience. In *Bad News*, discussed below in Chapter 10, curated narrative artifacts are mounted in a larger installation experience as stories that are told conversationally by a live actor. Another prospect, which I expound upon in Section 5.3, is to feed the curated artifact back into the simulation that it recounts. For example, characters in a simulated storyworld could tell stories about that world that are enabled by curation, and the very telling of those stories could alter the world, leading to new kinds of stories, and so forth. Indeed, a number of exciting possibilities emerge when the critical mounting step is reached.

### Aesthetic Posturing

As I write this dissertation, Google artist in residence Mario Klingemann is working to build neural networks that can generate human portraits in the style of the old masters. Writing about the project for Fast Company's *Co.Design* blog, Katharine Schwab notes, “The results of his attempts are often humorous, showing just how bad algorithms are at conjuring something as complex as a human face” [schwab2018ai]. In my opinion, however, the outputs that are dismissed as humorously bad are actually more compelling than their strictly representational counterparts. Figure 4.6 shows examples of outputs that cannot be easily distinguished from the works that the model was trained on, as well as ones that eschew representational convention in visceral expressions of quirks in the underlying model. Even if one enjoys representational visual art, which I admit I tend not to, I maintain that the outputs that are indiscernible from the genre works they mimic lack conceptual intrigue. While Klingemann's preferred outputs conceal their procedural origin, it surfaces in the other examples through curious artifacts that reveal something about the underlying mechanisms.

[Figure omitted]

#### Toward a New Aesthetics of the Computational

What is the point of generating artifacts that perfectly resemble ones created by humans? A generator that does that has the conceptual intrigue of a photocopier or a printing press—it is a machine for generic mass production. Computer art resembling the old masters is a magic trick whose allure will quickly wear off. Moreover, there is already a massive accumulation of human works, and appreciation of these works is often rooted in their human origins: humans enjoy the expression in a human-authored work of the particular will and worldview of the human author. If there is a choice between human-authored works and perfect facsimiles generated by computers, I think humans will almost exclusively prefer the former. Computer-generated artifacts that embrace a new aesthetics of the computational, on the other hand, are distinct from human works not just conceptually but also in terms of form. By leaning into their quirky aesthetics, these artifacts offer features that human works cannot. [Note: I should mention that one reason for building such a machine is to deeply study a creative process. In this pattern of inquiry, a computer program that generates artifacts is viewed as computationalizing the human creative process for producing such artifacts. By inspecting the operation of the program, it may be possible to learn about the human process that it models. Additionally, the program may be developed to instantiate a particular existing theory of that creative process (from the arts, social sciences, or humanities, for instance)—this is called *operationalization* [bagozzi1982representing]—in which case the resulting computer program may also be seen as rigorously testing the given theory. While I think it can be valuable to operationalize outside theories as part of the process of media creation, I am personally skeptical that doing so can reveal the creative process (or even tell us much about the theory). This is primarily for the reason that it is easy to create generators that produce the same artifacts but that are not *isomorphic* with regard to their underlying processes. This means that a computer program that generates identically to a human may do so in a way that does not resemble the human's process. Moreover, it is possible (or inevitable) to operationalize theories such that the computer program's instantiation of that theory is not isomorphic with the actual theory, which means the program in fact tests the theory constituted by itself, as opposed to the theory that it was developed to test. Even if isomorphism obtains, it may due to a hyperassociative mapping scheme between source and target structure [lenat1984why]. However, contemporary philosophers might argue that isomorphism is actually not important here, since the theory could be *multiply realized* (realized in different ways) in its original context (scholarly writing, likely) and in the computational one (both being viewed as inscription technologies) [kim1992multiple]. Alternatively, others argue that even perfect isomorphism is not sufficient for meaningful equivalence [blackmon2013searle]. These are murky issues in the philosophy of science [suppe1977structure,van2010scientific] and computational creativity [jordanous2012standardised], but thankfully the idea of operationalization as a mode of empirical investigation is not relevant in this dissertation: herein, I care about media creation and media artifacts, not the cognitive processes that drive creation or interpretation. That being said, I am still interested in two kinds of operationalization: the intuitive modeling of ideas from the arts and sciences and of tropes found in media, and the somewhat informal coevolution of a theoretical framework and a computational architecture, as articulated by Peter Mawhorter in his dissertation [mawhorter2016artificial, pp. 16–18] (through the influence of Phil Agre [agre1997toward]). As Part II of this dissertation will show, I make use of the former approach and the very subject of this paper, curationist emergent narrative, is a framework that has been derived by the latter method.]

#### Eschewing Conventional Narrative Aesthetics

Analogously to the case I have just outlined, there is a tendency among practitioners of procedural narrative to target the aesthetic qualities of conventional narrative forms, but this aesthetic posturing can work to suppress some of the intriguing qualities that are uniquely germane to computational works. As Espen Aarseth writes in his seminal *Cybertext*:

> To achieve interesting and worthwhile computer-generated literature, it is necessary to dispose of the poetics of narrative literature and to use the computer's potential for combination and world simulation in order to develop new genres that can be valued and used in their own terms. [aarseth1997cybertext, p. 141]

In emergent narrative, combination and world simulation are on full display, and this yields aesthetic qualities that diverge from conventional narrative works. In particular, the simulated storyworlds found in these systems work differently than the fictional worlds that are implied by conventional narrative works. While conventional works that are enabled by extensive worldbuilding efforts may feature a high level of detail in the terms of storyworlds that contextualize the stories recounted in the works, projects in emergent narrative go much farther by explicitly modeling the state of a storyworld and the mechanisms by which it may evolve. In making such commitments, quirks rooted in particular design emphases or simulation abstractions may manifest as bizarre quirks in the generated surface material. When this happens, the system creator can work to rectify the underlying cause so that the quirk will no longer manifest in the surface material, but this becomes a Sisyphean nightmare—it is (probably) not possible to simulate a world that works just like our world, or even like the implied storyworld of any work of fiction. Moreover, in ironing out these wrinkles, the generated material loses the distinctive qualities that are uniquely present in works of emergent narrative. When this is done, to call back to Aarseth's message quoted above, the practitioner fails to value this new genre on its own terms.

#### *Tale-Spin 's Secret Quantum Aesthetics*
Rather than viewing the idiosyncrasies of emergent narrative as wrinkles that should be ironed out, a practitioner might consider them instead as aesthetic qualities that should be amplified. In other words: why not lean into it? Let us consider *Tale-Spin* as a cautionary tale. As Noah Wardrip-Fruin argues in his chapter on the system in *Expressive Processing* [wardrip2009expressive], everything in its simulation hinges on character plans:

> But *Tale-Spin* does not begin with a complete virtual world that characters can move across. Instead, movement happens only when required by plans, and the world is fleshed out only to the degree required by plans. Similarly, *Tale-Spin* does not generate a complete set of interpersonal relationships when characters are created but rather only fleshes out the connections between characters that are required by plans. [wardrip2009expressive, p. 143]

 The inclusion of character planning is certainly not unusual among narrative works, but in *Tale-Spin* its presence is amplified and warped through the mechanisms of its detailed simulation. In a typical narrative work, a character plans by considering at most a few candidate routes of action before selecting one course to take. [Note: To be clear, here I am talking about narrative in a general sense, rather than computational narrative in particular, which means I likewise use ‘character planning' in a general sense, as opposed to automated planning.] As such, stories that work in this way are still about the actions characters do take, which is the stuff of conventional narrative. In *Tale-Spin*, however, characters may consider every possible action, all of which may be unavailable to them given the state of the storyworld, and in this way the simulation becomes as much about character *inaction*. Characters spin threads of *possible worlds* [kripke1963semantical,pavel1986fictional,ryan1991possible,ronen1994possible] in their minds, imagining what they *could* do, or what they would do, if the world were different. As Wardrip-Fruin puts it, by making it its operating principal, character planning in the system “creates the profusion of imagined worlds that define *Tale-Spin*'s fictions” [wardrip2009expressive, p. 143].

Beyond characterizing its simulation, *Tale-Spin*'s strange enumeration of plausible alternate realities yields the core aesthetic intrigue of its narrative material. Stories are typically about paths that characters take, but the stories that emerge from *Tale-Spin*'s simulation are about paths that characters do not take, and in fact *cannot* take. Both aesthetically and conceptually, this evokes Jorge Luis Borges's exhaustive unraveling of the possible that characterizes works such as “The Library of Babel” and “The Garden of Forking Paths” [borges1964labyrinths]. It is a quantum aesthetics, or more precisely, an aesthetics of the quantum.

Frequently, *Tale-Spin* stories are tragic, but they are structured like strange *proofs of tragedy*: each is an exhaustive, well-structured articulation of all the preferable realities that are cordoned off from a character due to the cruel upshot of initial simulation conditions. [Note: Of course, *Tale-Spin* is capable of producing stories that are not tragic, but these seem to be less interesting and less frequently discussed. In one of his papers on the system, Meehan notes, “I have observed that people who run the program usually make the problem *very* hard; they find the resulting ‘Trials and Tribulations' story more ‘interesting' (their word) than ‘Sweetness and Light' stories where all the characters like each other and do favors at the least suggestion” [meehan1977tale, p. 96].] Wardrip-Fruin, for example, recounts the tale of a hungry bear, Arthur, whose quest for honey was doomed from the start:

> And as we remember from the setup for the story, no one knows about any bear food in this world. Between Arthur and George they only know about one worm. So the story could not have turned out any other way. As soon as the audience decided that hunger was Arthur's problem, he was doomed. He made many plans, yet none of them had a chance of working. The end. [wardrip2009expressive, pp. 136–37]

 Given the initial state of the storyworld, there was no reachable later state in which Arthur Bear could acquire his sustenance, and the unfolding of the simulation simply proves this. The bear takes the only course of action available to him, and considers every unavailable course of action, and the resulting trace is an exhaustive and explicit proof of tragedy.

This is something strangely attractive about the alien narratology of a proof of tragedy. To be sure, it answers Aarseth's call, quoted above, “to dispose of the poetics of narrative literature and to use the computer's potential for combination and world simulation in order to develop new genres.” [Note: Unfortunately, as Wardrip-Fruin explains in *Expressive Processing*, Aarseth did not recognize *Tale-Spin*'s peculiarity as exemplifying an evocative new poetics [wardrip2009expressive, pp. 153–155]. This was likely due to a misapprehension of the system's underlying processes, which are heavily obscured in the surface text of its generated stories. This misapprehension is not unique to Aarseth's reading of *Tale-Spin*, but rather has been characteristic of much of the humanistic writing on the system. More broadly, the phenomenon of surface artifacts obscuring the richness of underlying systems is at play in many systems beyond Meehan's—fittingly, though, Wardrip-Fruin calls this the *Tale-Spin effect*.] Unfortunately, however, the detailed operations of *Tale-Spin*'s mechanisms for character planning are elided in the surface text of its generated stories. Instead, its text generator *Mumble* recounts only those actions that characters actually take, rather than the rich mental simulation of actions that they would prefer to take. This is how storytelling works conventionally, but *Tale-Spin* narrative is not conventional: while character action is the stuff of conventional narrative, Meehan's worlds are structured around debilitating character inaction—the result of exhaustive mental consideration of all the promising actions that a character cannot take. [Note: It is worth noting that stories are also frequently about inaction, as Marie-Laure Ryan has compellingly argued: “narrative plots are layered entities, made up not only of a linear sequence of factual events, but also of the projections, wishes, plans, and interpretations produced by the characters as they reflect upon the world of which they are members. Insofar as they link events and states in a causal chain, these mental constructs present and structure a story, and may therefore be called ‘embedded narratives'{”} [ryan1986structure, p. 107]. What is wild about *Tale-Spin*, however, is the degree to which it models such embedded narratives: characters in fiction consider alternatives and project onto possible worlds, but they do not exhaustively consider *every possible world* with respect to a targeted goal. Thus, the unique aesthetic of *Tale-Spin* is its fiercely recursive *nesting model* of embedded narratives. It is all about what Gerald Prince has called the *disnarrated*: “events that do not happen but, nonetheless, are referred to” [prince1988disnarrated, p. 2]; unfortunately, while such events are referred to in the system's execution stack, they do not manifest in the generated prose.] As a result of this unfortunate mismatch between the organizing principals at play in the underlying simulation and at the surface, the generated tales look nothing like proofs of tragedy, and *Tale-Spin* thereby renounces its own narratology.

#### The Wrong Narratology

Why did Meehan make this move? That is, why did he build a simulation about character planning—and as a byproduct, the consideration of possible worlds—and then curate this simulated material into accounts of character action? First, it is important to acknowledge that Meehan was far more concerned with the procedures of his underlying simulation than with the process of curating simulated material into a compelling narrative artifact. As I mentioned in the last section, *Tale-Spin*'s natural language generator, *Mumble*, was hastily developed in a single day. So one account of the mismatch between simulated material and surface narration is that *Mumble* is an incomplete system that was hacked together quickly, and thus does not represent an earnest attempt by Meehan to generate compelling narration. While this is almost certainly a contributing factor, I think it is still worth considering the particular way in which *Mumble* deviates from the narratological directives of *Tale-Spin*'s underlying simulation.

The classic *Tale-Spin* storyworld is inspired by Aesop's fables, and I think that *Mumble* was designed to curate its simulated material into narrative artifacts that read like those stories. Thus, instead of a surreal account of Wilma unraveling a tangle of possible futures, we get sentences like “Wilma was inclined to lie to Tom” (see Wardrip-Fruin's illustrative account of this ellipsis [wardrip2009expressive, pp. 144–145]). But *Tale-Spin*'s narratology does not work like Aesop's narratology—it is not even about character activity, let alone the moral ramifications or aphoristic suggestions of character activity—and so in targeting Aesop's style the system's generated tales fail to capture the interesting qualities of the underlying simulation. Moreover, because this underlying simulation does not work according to the principals that frame the narratology of Aesop, curation of *Tale-Spin*'s simulated material in the style of Aesop will tend to yield bad outputs. With a lot of work, *Tale-Spin*'s initial conditions can be rigged such that an actual Aesop tale is essentially reproduced in the *Mumble* generated telling—Meehan did this with “The Fox and the Crow” [meehan1977tale, p. 97]—but this does not mean that typical *Tale-Spin* runs will produce new fables that work like Aesop's.

More typically, the raw narrative material produced by the system, evoking a Borgesian hypersensitivity to the possible, loses much when it collapsed onto the rigid plane of actual executed actions. As such, the system's outputs have been widely ridiculed as boring stories, if not positively non-narrative. Indeed, Wardrip-Fruin has argued that the possible-worlds model implicit in its underlying processes (but missing in its surface) is perhaps the system's strongest connection to conventional modes of literature [wardrip2009expressive, p. 142]. More charitably, we might interpret Meehan's project differently by assuming that his primary aim was to operationalize the narratology of Aesop's fable (in the sense outlined in footnote Section 115 of this chapter), but this does not appear to have been the case [sack2013machine]. If it was, moreover, Meehan's planning-based approach is like the *Rube Goldberg machine* [olsen2017narrative] equivalent of a Proppian morphology [propp2010morphology]. Instead, we find the explicitly stated goal of modeling people [meehan1976metanovel, p. 4], though it is not clear whether this refers to human actors or human storytellers or both. In any event, as Wardrip-Fruin notes [wardrip2009expressive, p. 151], its particular model of human cognition has since become dated. Wardrip-Fruin, however, argues that this makes the system even more interesting, since its narratology is a strange one rooted in the particular time and place of the Yale AI Project of the 1970s [schank1978inference,schank1981inside,dehn1982nlp,slade1987yale]. If only this aesthetic intrigue was reflected in the surface prose of its generated artifacts.

#### To One's Own Simulation Be True

The curation procedure of *Tale-Spin*, as instantiated in the operation of *Mumble*, serves as a cautionary tale about aesthetic posturing in emergent narrative. At the level of curation, works of emergent narrative should be true to their underlying simulations, so that surface artifacts may actually reflect the interesting features of the raw material out of which they are composed. Furthermore, practitioners should avoid curation modes that reflect different narratologies than those that are operational in their simulation procedures. This may confuse the notion of artifact evaluation with regard to such systems, but I believe that is the cost of doing business in a new medium.

#### Lovable Quirks and Maddening Flaws

Lastly, I would like to acknowledge that this section may appear to contradict the ones that have preceded it in this chapter. That is, in earlier sections I identified a set of undesirable project characteristics that should be rectified, while in this section I argue for the celebration of quirkiness. As an example, one could view the tendency of Klein's murder-mystery generator to reveal the killer halfway through its generated stories as a quirk on the order of *Tale-Spin*'s deep dive into character planning. The difference, critically, is that one of these features supports the goals of a larger project, while the other does not. Klein is attempting to model the structure and tropes of a particular genre, the murder mystery, and in that genre the killer is not revealed at the time that the murder is first recounted. Thus, the revelation of such information at the wrong time may be viewed as an error with respect to the goals of the project. *Tale-Spin*'s obsession with character planning, on the other hand, is in line with Meehan's goals for the project—in fact, this obsession *is* the goal of the project. Moreover, the quirkiness of *Tale-Spin*'s rummaging in possible worlds is evocative, while Klein's system's premature revelation simply detracts from the intrigue of its generated artifacts.

Of course, this is just my opinion, and there is no objective distinction between quirks that are compelling and ones that are not. Judgments will vary according to audience tastes, and also according to that of creators. In the end, it is up to a practitioner to decide what is best for her project, but I hope my argumentation in this section has succeeded in repositioning certain characteristics of emergent narrative (and generative art, more broadly) that tend to be viewed as negative as being potentially, or even inevitably, positive. I would like to see more generative art, but like Aarseth I think the genre will only find its place when the works that exemplify it begin to embrace, and conventionalize, new aesthetic principals.