Chapter 25: Geometric Game Aesthetics — Playable Space and Procedural Beauty
RUNNING EXAMPLE — Sam’s Second Job
Sam, whom we met in Chapter 22 as a film editor studying pacing, has taken on a consulting contract for a mid-budget studio developing a roguelike deck-builder. The studio is not asking Sam about pacing in the cinematic sense. It is asking whether the game is fun, which the producer has told Sam, with a straight face, is “what we’re paying you to figure out.” The game has a playtester panel whose reactions range from “I lost sleep to this” to “I stopped after ten minutes because nothing was happening.” The studio wants to ship. Sam has begun to notice that the players who loved the game and the players who bounced are describing, with different affect, a recognizably similar object — the same mechanics, the same art, the same sound design. What differs is the trajectory each took through the game’s possibility space: the loved-it players threaded curves through the state manifold that their fellow testers never opened. Sam is trying to describe this to the producer in a way that is actionable rather than mystical. This chapter is for people in Sam’s position.
We close Part V with the modality whose pipeline analog is the least obvious and whose theoretical payoff is, we argue, the most distinctive. A painting is a static object whose embedding cloud is a single multiset in latent space (Chapter 23). A building is a static configuration traversed along a (largely) architect-specified promenade (Chapter 24). A book is a paragraph sequence read at the reader’s pace; a piece of music is a token sequence heard at the work’s pace. A game is something else. A game is a possibility space whose actualization is co-authored by the player in real time. The aesthetic object is not the game’s art, not its mechanics, not even its rule system — it is the trajectory the player carves through the state space during a given session. The art, mechanics, and rules are the scaffolding that makes certain trajectories more likely than others. This is the framework’s central commitment for this modality, and it leads to a specific, falsifiable claim about where the aesthetic signal should be largest.
25.1 What the Embedding-Cloud Analog Even Is
The first problem in applying the framework to games is that the “unit” is not clear. Is the unit the rule set (a single object, like a painting), the session (a trajectory through the rule set, like a book’s paragraph sequence), the aggregate playerbase trajectory (a cloud of trajectories), or the playable space itself (a geometric object whose intrinsic structure matters before any player touches it)? The framework needs a choice.
Definition 25.1 (Game State Embedding). Let $G$ be a game with state space $S$ (the set of all reachable configurations of the game’s internal variables), action space $A$, and transition function $T: S \times A \to S$. A game-state encoder $\phi: S \to \mathbb{R}^d$ maps each state to a $d$-dimensional embedding. Candidate encoders include (i) a transformer over the serialized game state (for discrete-state games such as card games, roguelikes, and chess variants), (ii) a convolutional or vision-transformer encoder of a canonical rendered screenshot (for real-time games), and (iii) a multi-modal encoder combining game-state text, rendered image, and audio snippet (for rich games).
Definition 25.2 (Session Trajectory). A session of $G$ is a finite sequence of states $s_0, s_1, \dots, s_k \in S$ generated by a player’s actions. The session trajectory in embedding space is $\phi(s_0), \phi(s_1), \dots, \phi(s_k) \in \mathbb{R}^d$.
Definition 25.3 (Game Cloud). The game cloud of $G$ is the empirical distribution of embeddings over a large sample of sessions from a reference player population: $C(G) = \{\phi(s) : s \text{ appears in some session of } G\}$. The game manifold $M_{\text{game}}$ is the corpus-level geometry across a reference set of games.
With these definitions, the four channels of Chapter 17 have analogs, but they are allocated across session and game levels rather than, as in the book case, across paragraph and book levels.
The Four Channels, Instantiated for Games
A. Divergence of $C(G)$ from a game-corpus prior measures the game’s typological distinctiveness — how much its state-space coverage departs from the coverage of comparable games in the reference corpus. A roguelike whose runs cover regions of state space that other roguelikes do not should score high on this channel.
B. Internal coherence of $C(G)$ — pair_sim_mean over sampled states — measures how self-similar the game’s state space is. A game whose states all look alike (many idle games, most slot-machine-style mobile monetizers) will have high pair_sim_mean and low internal variation. A game whose states span very different visual and structural registers (an RPG moving between dungeons, overworlds, cities, and combat screens) will have lower pair_sim_mean. We predict, for reasons discussed in §25.3, that the sign of this channel against aesthetic rating is negative for games — the opposite of books and the same as music. Games reward contrast.
C. Trajectory geometry — step_mean, step_std, recur_rate, curvature, path_eff — is computed over sessions. Each session contributes its own trajectory through $C(G)$; the aggregate statistics over the playerbase are the game’s trajectory signature. This is the channel we claim is distinctive to the modality.
D. Lasso-latent directions in a PCA of $C(G)$ should recover interpretable game-design features: difficulty ramp, resource-accumulation tempo, exploration-versus-exploitation balance, narrative density, visual-chapter pacing.
25.2 The Rating Analog
Games have better rating-analog candidates than painting or architecture. Steam reviews (binary thumbs with free text), Metacritic scores, average session length, 30-day retention, completion rate, and within-game progression telemetry are all plausible targets. None is perfect: Metacritic aggregates professional reviews that are known to be heavily confounded by marketing spend and release timing; retention is confounded by monetization design; completion rate is strongly genre-dependent. But unlike painting, we have millions of rating events per title, and unlike architecture, the rating events are digital-native and time-stamped.
The cleanest rating-analog for framework purposes is almost certainly session-level enjoyment inferred from within-session behavior — a measure derivable from telemetry that correlates with but does not equal review score. This is, in effect, the game-industry analog of eye-tracking data for painting: it is a measure of how the work is actually received, moment by moment, rather than how it is summarized after the fact. Several major publishers have this data; academic access remains limited. This is the empirical gap we flag for the chapter.
25.3 The Central Prediction: Trajectory Dominates
The framework makes one specific prediction for games that distinguishes this modality from every other application in Part V.
Prediction 25.1 (Trajectory Channel Dominance). For games, the trajectory channel (C) contributes more of the residualized aesthetic signal than it does for books, music, painting, or architecture. Formally, let $R_A, R_B, R_C, R_D$ denote the within-genre residualized Pearson correlations of the four channels with an aesthetic target. For books (Chapter 17, §17.4), the dominant channels are A (divergence) and B (coherence), with C weaker; for music (§17.6), C is intermediate; for games, we predict $R_C > R_A, R_C > R_B, R_C > R_D$.
The reasoning is structural, not speculative. For a book, the trajectory is fixed by the author — the reader reads the paragraphs in the order the author wrote them. For music, the trajectory is fixed by the composer. For a painting, the trajectory is semi-free (the eye may saccade variably) but the static object constrains it. For a building, the trajectory is constrained by the circulation plan. For a game, the trajectory is the player’s principal creative contribution to the work. The aesthetic object is the trajectory. The game is the constraint set under which trajectories are authored.
If Prediction 25.1 is correct, then game criticism’s long-standing difficulty — articulating what makes a game good in terms that criticism of other media would recognize — is not a failure of game criticism. It is a consequence of the fact that game aesthetics live predominantly on a channel that other media’s aesthetics do not live on. Literary critics can talk about paragraph-level coherence (channel B) because that is where the literary signal lives; game critics who try to do the same for games are looking at the wrong channel.
Prediction 25.2 (Aesthetic Feel as Trajectory Curvature × Step Variance). Within the trajectory channel, the specific sub-features that carry the most signal for games are curvature (how sharply the trajectory bends, in the discrete-state sense of $\|s_{t+1} - 2 s_t + s_{t-1}\|$ in embedding space) and step variance (step_std, the variability of the size of successive state changes). We predict that the product $\text{curvature} \times \text{step\_std}$, computed over a session trajectory and averaged over a playerbase, is the single strongest predictor of within-genre residualized enjoyment. A session that bends and surprises is felt as alive. A session whose curvature is zero (monotonic grind) is felt as work. A session whose step variance is zero (every action feels the same) is felt as empty.
This is the chapter’s sharpest claim. It is risky. It is also falsifiable with existing telemetry data at any major studio.
25.4 Connection to Existing Game-Theoretic Literature
The predictions above are in dialogue with four traditions.
Csikszentmihalyi’s flow. Mihaly Csikszentmihalyi’s flow theory (Flow: The Psychology of Optimal Experience, 1990) posits that optimal experience occurs when challenge and skill are matched within a narrow band — the flow channel — and that outside this band, experience becomes anxious (challenge too high) or boring (challenge too low). In the framework’s language, flow is a constraint on the trajectory’s curvature: the session’s step sizes should be just above the player’s capacity to predict the next state. Curvature that is too high (a sharp spike of difficulty) throws the player into anxiety; curvature that is too low (a flat grind) produces boredom. Prediction 25.2 is, in effect, a framework-native restatement of Csikszentmihalyi, with the specific operationalization that flow is a geometric property of the session trajectory rather than a psychological property that the game only indirectly produces.
Koster’s A Theory of Fun (2004). Raph Koster’s influential short book argues that fun is pattern-learning: a game is fun while the player is still identifying and mastering its underlying patterns, and it becomes tedious once the patterns are fully learned. In the framework’s language, Koster’s claim is that the aesthetic signal is concentrated in the trajectory’s novelty — the rate at which the session visits regions of state space the player’s internal model has not yet assimilated. A game whose trajectory stays within a shrinking novelty envelope (trajectory curvature → 0 as sessions accumulate) is a game that has been mastered. Koster’s “learning curve” is a curvature curve on the session trajectory. Prediction 25.2 operationalizes Koster in a way that does not require direct access to the player’s internal model — only the geometric properties of the trajectory itself.
Procedural-content-generation (PCG) literature. Research on PCG — Smith et al.’s search-based methods, the Game AI Research Network’s benchmarks, Shaker, Togelius, and Nelson’s synthesis — has produced a substantial literature on measuring generated content. The quality measures most commonly used (fun, challenge, balance, surprise) have tended to be bespoke and hand-crafted per game. The framework proposes a modality-general alternative: rather than defining game-specific measures, evaluate generated content by the expected geometric properties of the sessions it induces. A PCG system whose outputs produce trajectories with high expected curvature × step variance is generating content that is, under the framework’s prediction, aesthetically stronger. This is a specific and testable bridge between PCG research and the framework.
Juul’s The Art of Failure (2013). Jesper Juul’s observation that players seek out failure in games — that failure is not the opposite of aesthetic success but one of its ingredients — becomes, in the framework, a claim about acceptable trajectory discontinuities. A trajectory that terminates in failure and restarts has high curvature at the failure point; players tolerate this curvature because it is the content. Juul’s empirical observation is that excessive curvature (repeated failure with no progress) and zero curvature (no failure, no challenge) are both aesthetically inferior to moderated curvature. The curvature × step_std quantity of Prediction 25.2 is, roughly, Juul’s sweet spot made geometric.
25.5 Genre, as Ever
The within-genre residualization that destroyed 85% of the books’ $R^2$ (Chapter 17, §17.3) and 91% of the music’s (§17.5) will be at least as severe for games. Genre in games (roguelike, Soulslike, 4X strategy, farming sim, walking simulator, puzzle, rhythm, battle royale) is more strongly visually and structurally encoded than either book or music genre; a game’s state-space embedding cloud will identify its genre with essentially perfect accuracy. Any naive correlation between trajectory features and enjoyment will be dominated by genre.
Prediction 25.3. The honest, within-genre residualized correlation between curvature × step_std and session enjoyment will be on the order of $R = 0.10$–$0.20$, consistent with the residualized book and music results (Chapter 17). Unresidualized correlations will be much larger (we would not be surprised by $R \geq 0.4$) and will be reporting genre. We anticipate practitioner and industry claims of very large effect sizes and flag them, in advance, as probable genre confound.
25.6 Procedural Beauty: The Game as Generator
A further and more speculative prediction. The foregoing treats the game as a fixed rule set and asks about the trajectories it produces. A different framing asks about the game as a generator — about the set $\{C(s) : s \in S\}$ of embedding clouds induced by every possible game session, not just the ones players actually play.
Definition 25.4 (Playable-Space Manifold). The playable-space manifold $P(G)$ of a game $G$ is the set of reachable state embeddings $\{\phi(s) : s \text{ is reachable in } G\}$ together with the transition geometry $\{(\phi(s), \phi(T(s, a))) : s \in S, a \in A\}$. This is a directed graph in $\mathbb{R}^d$; it encodes the possibility space of the game independent of any specific player.
Prediction 25.4 (Procedural Beauty as Playable-Space Richness). A game’s long-run aesthetic standing — the extent to which the game remains played, discussed, and valued years after release — correlates with geometric properties of $P(G)$ that are independent of any specific playthrough. Specifically: the volume of $P(G)$ (how much of embedding space the playable states cover), the connectivity of $P(G)$ (how well different regions are reachable from one another), and the granularity of $P(G)$ (the distribution of step sizes available from typical states). A game whose $P(G)$ is geometrically rich supports many aesthetically strong session trajectories; a game whose $P(G)$ is geometrically thin supports few.
This is the framework’s explanation for why some games — Dwarf Fortress, Minecraft, Chess, Go, Civilization — retain aesthetic interest over decades while other, similarly-well-made games do not. The retained games have playable-space manifolds whose geometric richness continues to reward new trajectories long after the more-common manifolds have been exhausted.
We do not have direct empirical support for Prediction 25.4. We note it because it is the cleanest framework-native statement of what game designers mean when they say, of a long-lived game, that it has depth.
25.7 What Would Refute the Framework in Games
The framework is refuted in games if any of the following is observed under a properly-designed study (within-genre residualization, player-disjoint folds, pre-registered feature set, $n \geq 1{,}000$ games spanning at least five genres, with telemetry-derived enjoyment targets).
First, if the trajectory channel is not dominant — if $R_C$ does not exceed $R_A$, $R_B$, and $R_D$ after within-genre residualization — Prediction 25.1 fails. The framework’s central and most distinctive claim for this modality is falsified. The framework might still recover useful per-channel signal, but the argument that games live on a different aesthetic channel from other media would collapse.
Second, if curvature × step_std does not outperform alternative trajectory features (or outperforms only trivially) as a predictor of within-session enjoyment, Prediction 25.2 fails. The specific geometric operationalization of flow would then need to be replaced by a different one, possibly retaining the general trajectory-dominance claim.
Third, if the sign of the coherence channel matches books rather than music — if games, like books, reward internal coherence — then the modality taxonomy of §17.6 requires a different partition than we have proposed, and games belong on the continuity-rewarding side rather than the contrast-rewarding side. We think this unlikely (the empirical folklore that the most-loved games are those with the greatest variety of experience speaks against it), but we note it as a clean falsification condition.
Fourth, if procedural-beauty measures on $P(G)$ (Prediction 25.4) have no correlation with long-run game standing after controlling for budget, marketing, and genre — if geometric richness of the playable space does not distinguish Dwarf Fortress from its less-durable contemporaries — then the long-run aesthetic claim fails, and games’ durability is being explained by factors orthogonal to playable-space geometry.
25.8 The Empirical Gap, Stated Plainly
We have not done this work. We do not have access to the session-telemetry of a major publisher, and the publicly available game datasets (the VGDL corpus, the General Video Game AI benchmarks, Steam review data) do not yet come with both game-state embeddings and fine-grained enjoyment signals on the scale the book and music studies had. The pipeline of §25.1 can be built; the data can, in principle, be obtained; the study is not, at this writing, done. We state this as plainly as we stated it for painting (§23.6) and architecture (§24.8), and for the same reason: the book’s credibility depends on its distinction between confirmed findings (Chapter 17, the books and music results) and framework predictions (this chapter and the two preceding).
What we claim, narrowly, is that the predictions of §§25.3, 25.4, and 25.6 are specific commitments of the framework applied to games, and that they are falsifiable with data that exists in principle. A reader who is skeptical should not, for these chapters, ask “is this shown?” The answer is no. The reader should ask “is this the framework’s specific, sharp, advance-stated prediction for this modality, or is it decoration?” We have tried to make sure it is the former.
25.9 Sam, Reconsidered
Return to Sam and the deck-builder. The studio’s question — “is the game fun?” — is not answerable as asked. The framework-native restatement is: for the distribution of sessions the game’s current design induces, what is the average curvature × step_std on the residualized enjoyment axis, and where on the curvature–step_std plane does the current design sit relative to titles of comparable genre that succeeded? This is a question the studio’s telemetry team can actually compute. The producer will push back that this is not what players mean by fun. The producer is correct that it is not exactly what players mean. The producer is incorrect that the distinction matters at the level of design decisions. What matters at that level is whether the game’s session trajectories occupy the geometric regime that historically correlates with successful reception in its genre, and if not, which channel is off and by how much.
This is modest. It does not promise to turn game design into a science. It does promise to turn one specific, widely contested question — “is this game fun” — into a question with a geometric answer that is computable from data the studio already collects. That is the chapter’s operational bottom line. The theoretical bottom line, and the thread running through this part of the book, is the one the framework has insisted on from Chapter 1: aesthetic judgment has structure; that structure is geometric; the geometry is modality-specific; and the central discovery of the work — that the sign of the coherence and trajectory channels flips between books and music — admits a clean extension to the modalities we have not yet measured. Games are, we predict, the modality on which the trajectory channel’s dominance is most extreme. If we are right, much of what the games industry has been trying to measure with scalar review scores and retention dashboards has been the wrong channel of the wrong geometry. If we are wrong, we will have been wrong in a way that admits clean diagnosis, which is what the framework was built to make possible.
25.10 Bridge to Chapter 26
We close Part V with a modality whose aesthetic object is the trajectory. Chapter 26 takes up AI curation — the algorithmic selection, ranking, and recommendation of aesthetic objects at scale — where the aesthetic object under judgment is not a trajectory but a distribution: the distribution of works a recommendation system chooses to place in front of a user. The geometry shifts again. But the framework’s core commitments — the four channels, the directionality of the coherence and trajectory signs, the centrality of within-genre residualization, the distinction between confirmed findings and framework predictions — carry through. Sam, Hiroshi, Leona, and the novelist Maya from Chapter 20 are all, in the end, making decisions on the same manifold. The last chapters of Part V, and the concluding Part VI, follow the manifold through.
For the empirical grounding of the trajectory-channel claim — the $\text{step\_mean}$ sign flip between books ($\rho = -0.096$) and music ($\rho = +0.071$) reported in Chapter 17, §17.6 — see especially the cross-modality table in that chapter. For the Csikszentmihalyi flow restatement, see the general treatment of optimal-search aesthetic dynamics in Chapter 11. For the procedural-beauty claim of §25.6, the nearest mathematical treatment is in Chapter 10 on parallel transport and style holonomy.