The Grammar of Simulation

Abstract

Simulation games are built from a shared but undocumented vocabulary of rule-systems. Naming it shows designers which combinations are crowded and which are empty, what each part costs in computation, writing, and clarity, and where new techniques could change a game. This report makes that vocabulary explicit and machine-readable, then asks how mathematical, AI, ML, and LLM results from 2019 to 2026 change it. We decompose 179 simulation games (1968 to 2026, 31 subgenres) into 199 atomic mechanics, analyze each mechanic on its own (definition, levers, failure modes, dynamics, mathematical basis, and six anchored scores), map 46 techniques onto mechanics through 223 typed effect edges, and synthesize 42 design opportunities, each with a validation experiment and a dated, falsifiable prediction. All results come from an open, versioned dataset in which every entity cites its sources, and every sentence that interprets the data is checked against it on every build.

Findings.

1. Simulation games are compositions of six layers: substrate, dynamics, coupling, control, legibility, and pacing. Clustering games by the mechanics they use, with the number of clusters chosen by resampling stability rather than by eye, recovers design families (minds, managed crowds, planned economies, build-and-test, hands-on work, authored decisions, generated worlds) that cut across market subgenres; 20 of 24 clusters mix subgenres, and 12 survive resampling.
2. Later games are built mostly from old parts. 69 percent of mechanics appear in a catalog game released in or before 2000, but random sets of the same number of games from any period cover nearly as many (p = 0.14), so that share mostly reflects how widely mechanics are shared rather than early invention. Reuse is the finding that holds: games released since 2010 draw 84 percent of their mechanic uses from mechanics already seen in or before 2000. 33 mechanics first appear in 2010 or later, and 28 percent of possible mechanic pairs have ever co-occurred.
3. Deeper mechanics are rated harder to read. On rubric scores from a single model rater, of which an independent second rater has so far scored 60 of 1194 (5 percent), legibility falls as emergence potential rises (ρ = -0.42) and as compute cost rises (ρ = -0.35); emergence rises with tuning difficulty (ρ = 0.61).
4. AI techniques target what games show and say, not what they are built from. Exposure concentrates in narrative, control, agent, and physics mechanics. 99 mechanics, carrying 52 percent of all mechanic uses in the catalog, are affected by no technique in the set. Explanation is the rarest effect a technique has on a mechanic (6 of 223 edges).
5. What shipped mostly stays at the edge of simulation state. Neural rendering, asset and motion generation, and language characters for dialogue and companionship have shipped. The shipped techniques that write into game state are narrow: cached model outputs as crafting rules, language-model characters in a few titles, and learned or search-based opponents. Frame-generating world models and learned physics remain prototypes or research. Of the 35 techniques that can run during play, 10 break reproducibility outright.
6. The best near-term opportunities keep the simulation symbolic and let learned models read, explain, decorate, and propose through validated, logged interfaces. The highest-priority opportunity is Why-query over the simulation log.

The dataset, schemas, contract, and every projection are open for extension; the procedure for contributors, human or agent, is in AGENTS.md.

1. Introduction

A simulation game is a game whose state changes by rules the player did not author and does not fully control. The player builds a road and the traffic is the city's reply; the player sets a tax and the citizens decide whether to stay. Everything distinctive about the form follows from that gap between intent and outcome: the pleasure of watching a system respond, the frustration of not understanding why it did, and the stories players tell about runs no designer scripted.

The genre is old and large. Its lineage runs from text programs on mainframes (a grain-and-population ledger in Hamurabi, a wagon party's rations in The Oregon Trail) through SimCity and The Sims to Dwarf Fortress, Factorio, and EVE Online. Its methods are less visible than its titles. Designers share a working vocabulary (needs meters, production chains, storytellers, overlays) but no common, inspectable map of which rule-systems exist, which games combine them, how each one fails, and which are near their ceilings.

That map matters now because the ceilings are moving. For forty years the depth of a simulation game was bounded by a few constraints: how much compute a consumer machine could spend per tick, how much content a studio could author, how much hidden state a player could understand, how tightly a multiplayer game had to stay deterministic, and how believable a rule-driven agent could be. Results from 2019 to 2026 in learned world models, language-model agents, learned physics, accelerated simulation, automated design, and on-device inference press on each of those constraints, unevenly and with new costs.

This report does four things:

1. Catalogs 179 simulation games across 31 subgenres, from 1968 to 2026, and decomposes each into the mechanics it uses (2023 typed game-to-mechanic edges, each marked core or supporting, with implementation notes for core mechanics).
2. Analyzes 199 mechanics one at a time: rule-level definition, player experience, design levers, failure modes, dynamics, mathematical basis, and six anchored scores (legibility, emergence potential, authoring cost, compute cost, tuning difficulty, AI leverage).
3. Maps 46 mathematical, AI, ML, and LLM techniques onto those mechanics through 223 typed affects edges that say how each technique changes each mechanic (replaces its substrate, deepens it, enables a new form, cuts its authoring cost, explains it, tunes it, or accelerates it).
4. Synthesizes 42 design opportunities, each with a validation experiment a small team could run and a falsifiable prediction with a resolve-by date.

Every number in this document is rendered from the dataset at build time; the dataset hash on the title page identifies the exact version. The report, the wiki, the tables, and the graph are projections of the same YAML files, so a correction to one entity updates all of them.

2. Method

2.1 Unit of analysis: the mechanic

A mechanic here is a rule-system that passes four admission tests (CONTRACT.md, I12):

• Nameable. Designers would recognize it in games other than the one being examined.
• Atomic. It owns one kind of state and one update rule. If it splits into two parts that each appear without the other in shipped games, it is two mechanics.
• Tunable. It has at least two levers.
• Removable. A game could drop it and keep running.

The tests exclude two common confusions. Dynamics such as "cascading failure" or "death spiral" are outcomes of coupled mechanics, so they appear in mechanics' dynamics fields rather than as entities. Genre-level patterns such as "colony management" are combinations, so they appear as game decompositions or opportunities. The result sits between the MDA framework's mechanics layer and a designer's pattern language: concrete enough to implement, general enough to recur.

Each mechanic belongs to one of 21 families, assigned by the state it owns rather than the genre it appears in: Weather is environment, Labor market is economy, Agent inspection is information.

2.2 Corpus

The game catalog was built to cover every subgenre in the vocabulary, every lineage-defining title from the 1960s onward, recent releases through 2026, research demonstrations that run as interactive simulations (such as Generative Agents (Smallville) and Project Sid), and works included for a single mechanic (marked boundary_case; 22 such works). The catalog is a purposive sample, not a census: prevalence figures describe this corpus, and the earliest-appearance year of a mechanic is a lower bound on its true origin.

2.3 Schema and scoring

Every categorical field draws from a closed vocabulary (vocab/), and every ordinal field names an anchored rubric (vocab/rubrics.yaml). Scorers match the anchor text and take the lower anchor when between two. This trades resolution for consistency across contributors: a 3 in legibility means "understood after deliberate attention and the game's info tools" whoever assigned it. The original scores were assigned by one model rater (claude-opus-5-5) matching the anchor text. Independent second ratings now cover 60 scores; fields with ratings store their lower median. Scores remain judgments, and the rubric for each is reproduced in the wiki.

Relations are typed edges with one owner each (vocab/edge_types.yaml). Inverses (which games use a mechanic, which techniques affect it) and all aggregates (prevalence, earliest appearance, clusters, technique exposure, opportunity priority) are computed by the build and never stored, so they cannot drift from the edges they summarize.

2.4 Evidence

Each entity carries an epistemic status. asserted means written from domain knowledge without a citation; sourced means at least one cited source supports its factual fields; verified requires a second, independent run. Techniques must always cite. At this version, 394 entities are sourced or verified and 71 remain asserted; the asserted ones form a public verification queue (dist/analysis/verification-queue.csv).

2.5 Production

The dataset was produced by parallel research runs, each recorded in log/runs/ with its operator, model, method, and the proposals it could not act on. Parallel runs partitioned work by file, not by topic: taxonomy runs analyzed mechanics; catalog runs verified release facts against public sources and decomposed games; technique runs searched the 2019 to 2026 literature and developer announcements. Merge-time runs then edited entities across partitions: the synthesis run added symmetric synergy and tension edges in one pass, and a review run added mechanic tags to games after an independent review found widely shared mechanics (pause controls, placement, tech trees) under-tagged, recoded several decompositions, and tightened every prediction. Each cross-partition edit is attributed in the entity's provenance.reviewed list. A validator enforcing the rules in CONTRACT.md §5 gated every merge. The same procedure (AGENTS.md) is what a new contributor, human or agent, follows.

2.6 Limitations

• Sampling. Commercial, English-language, PC-centric titles are over-represented; mobile management games and non-Western simulation traditions are under-represented.
• Scoring. Most rubric scores remain judgments by the original model (claude-opus-5-5), with an independently rated subset. They are not measurements of player behavior. Correlations among scores (§5.6) and between a score and the technique graph (§8.4) partly reflect that rater's internal consistency. Each score can carry ratings from further raters, each with a run and a rationale; the original score counts as one of them, and the stored value is their lower median (E037). A rater must come from a different operator, and a different model for agents (E035), and must first agree with a calibration set of 20 mechanics (E036). The calibration's reference scores are public, so it shows that a rater reads the anchors as the set does, not that the rater scored blind. 60 of the 1194 mechanic scores have a second, independent rating so far.
• Decomposition depth. A game's mechanic list records what is present and meaningful, not a complete specification. Absence of an edge is weak evidence of absence.
• Recency. Technique maturity and horizon are judged as of the run date and age quickly; each technique records the sources the judgment rests on.

3. The catalog

The corpus spans 1968 to 2026. Its oldest entries are text programs written for teaching (Hamurabi, The Oregon Trail, Lemonade Stand) and a zero-player automaton (Conway's Game of Life); its newest are releases from 2024 to 2026 and research demonstrations whose agents are language models.

Games in the dataset by decade of first release.

3.1 Lineages

Subgenre labels name design lineages, not markets. Prison Architect is a management game by lineage whose inmates carry colony-sim needs and moods; Kenshi is an RPG whose world runs on faction and economic simulation; The Legend of Zelda: Breath of the Wild and Noita are included as boundary cases because their material systems (Material interactions, Falling-sand materials) influenced simulation design beyond their genres. Section 5 shows that clustering games by the mechanics they actually use cuts across these labels.

3.2 What is simulated

3.3 How failure and story work

Generative narrative (a learned model producing story text or events at runtime) is still rare in the catalog and concentrated in recent AI-native titles:

3.4 Most emergent titles

Emergence and influence are rubric scores (vocab/rubrics.yaml): emergence 5 means players report histories no one could have authored; influence 5 means the game created a lineage or changed the medium.

4. A grammar of simulation games

4.1 Six layers

Decomposing the catalog suggests that every simulation game is assembled from six layers. A mechanic usually lives in one layer, and a game's character comes from which layers it invests in.

1. Substrate: what exists. State is held as aggregate stocks (a treasury, a population count), individual agents (a colonist with needs), fields on a grid (pollution, temperature, land value), networks (roads, pipes, social ties), or continuous bodies (a wing, a chassis). The substrate fixes the cost curve: aggregates are cheap and faceless, agents are expensive and legible, fields scale with map area, continuous physics scales with contacts and timestep. The substrate field on every mechanic records this (

). 2. Dynamics: how it changes. Update rules form feedback loops. Most simulation games pair a reinforcing growth loop (more citizens, more taxes, more buildings) with balancing loops that resist it (congestion, pollution, unrest, raids). Difficulty is the relative strength of those loops; a death spiral is a reinforcing loop running in reverse. The feedback field records the dominant polarity per mechanic. 3. Coupling: how subsystems touch. Depth comes less from the number of systems than from how many share state. Systems couple through shared stocks (power, money, labor), shared space (a factory lowers neighboring land value), and shared agents (the same colonist hauls, eats, and grieves). Dwarf Fortress is deep because nearly everything touches nearly everything else; a game with many isolated systems feels like many small games. 4. Control: how intent enters. The player acts directly (an avatar, a unit), places and zones, sets parameters, issues orders that agents pick up, writes programs, or speaks. Each step away from direct control widens the control gap, the designed distance between what the player wants and what happens. Too narrow and the game is a toy; too wide and the player is a spectator.

5. Legibility: how state leaves. Overlays, inspectors, tooltips, advisors, alerts, graphs, chronicles, and forecasts expose hidden state. This layer is what lets deep simulations ship: RimWorld itemizes every mood modifier, and Paradox games nest tooltips until each number decomposes into causes. Legibility mechanics own no simulation state of their own; they are the interface between coupling and comprehension.
6. Pacing and stakes: when the player looks and what can be lost. Time models (pause, speed, turns, wall clock), failure models (soft setbacks, permadeath, cascading collapse), and narrative directors decide when attention is demanded and what a mistake costs.

A compact description of a simulation game is its investment profile across these layers. City builders invest in substrate (fields and networks) and legibility (overlays); colony sims in agents, coupling, and pacing (storytellers); vehicle sims in continuous substrate and direct control; programming games in control alone, with nearly no hidden state. The clustering in §5 recovers these profiles from the data without being told the subgenres.

4.2 The families, one at a time

Each family below is described by the state it owns, its canonical mechanics, how it characteristically fails, and what it couples to. Tables list every mechanic in the family with how many catalog games use it and its rubric scores; each mechanic's full analysis (definition, levers, failure modes, dynamics, open questions) is on its wiki page.

Resources and production

Owns stocks of goods and the recipes that transform them. Production chains is the backbone of city builders, factory games, and colony sims alike; Storage and buffers, Spoilage, and Byproducts and waste decide whether a chain is a puzzle or a ledger. The characteristic failure is the spreadsheet: chains whose only decision is ratio arithmetic. The best implementations make chains spatial (Conveyor transport) or social (who does the work, and in what mood).

Logistics and networks

Owns movement across space or graphs: roads, belts, pipes, rails, transit lines, service reach. These mechanics turn production into geography, and they are where compute and legibility first bind: Traffic congestion is the most visible place a city builder can be wrong, and Walker service delivery made service coverage legible by putting it on walking bodies. Networks couple to almost everything, so their bugs (pathfinding loops, gridlock) spread to every system they feed.

Economy and finance

Owns money, prices, and valuation. Most games use only Treasury and budget and Taxation as a scorekeeping layer; fewer simulate markets with Supply-demand pricing, competitors (Rival firms), or player-made prices (Player-driven markets). The characteristic failure is the dominant strategy: a price or loan loophole that converts the economy into an exploit, which is why economies are the first target for automated balancing (§9).

Agents and behavior

Owns individual minds: needs, utilities, moods, memories, skills, schedules. Needs decay plus Utility action selection is the standard architecture since The Sims; Smart objects moved behavior knowledge into the world so new objects bring new behaviors. Its variety is paid for in animations, thoughts, and dialogue, which is why it is among the families most exposed to language-model techniques (§8).

Social relations

Owns ties between agents: relationship values, opinion modifiers, reputation, factions, gossip, romance, and conversation. Rule-based social systems are cheap but read as formulaic; Social exchanges in Prom Week and Versu showed richer social logic is possible at high authoring cost, and Natural-language interaction opened conversation as an input. The characteristic failure is the visible formula: players learn that one compliment always adds a fixed amount.

Population and lineage

Owns birth, death, migration, inheritance, and structure. Immigration by attractiveness is the city builder's growth engine; Population strata turns growth into progression; Trait inheritance and Dynasty succession stretch play across generations. These mechanics define the time horizon of a game more than its time model does.

Governance and policy

Owns rules about rules: policies, laws, elections, interest groups, satisfaction, unrest. Policy menus (Policy toggles) are universal; Policy influence network in Democracy 4 made the causal web itself the interface; Player legislation in Eco and A Tale in the Desert let players write rules for each other. The failure is the menu: a fixed list of authored options that no player can extend.

Space and construction

Owns layout: zoning, placement, adjacency, rooms, structure, terrain, and the scoring of player-built designs. Zoning is indirect construction; Building placement is direct; Procedural build assist sits between them, letting rough gestures produce detailed structure. Design performance rating closes the build-test loop in coaster, bridge, and machine builders.

Physics and materials

Owns continuous dynamics: rigid bodies, vehicles, flight, orbits, fluids, heat, deformation, destruction, material interactions. These are, on average, the most compute-bound mechanics in the taxonomy, and their legibility depends on how closely they match the physical intuitions players bring. The failure is the uncanny edge: a physics model accurate enough to invite intuition but wrong in the cases players probe.

Environment and ecology

Owns the world's own processes: weather, seasons, food webs, habitats, pollution, fire, contagion, water, climate. Environmental mechanics are where simulation games most often lean on real science (Contagion follows compartment models; Food web follows predator-prey dynamics) and where calibrated learned models can substitute for hand-tuned approximations.

Biology and health

Owns bodies: survival meters, injuries, diagnosis, creature care, husbandry. Survival meters define a subgenre; Injury body model tracks wounds per body part in colony sims; Creature care is the oldest bond between player and simulated being.

Time and pacing

Owns the clock. Pause and speed control is near-universal in real-time management games; Real-time clock ties play to the calendar; Daily action budget makes days a scarce resource; Crisis countdown turns the future into a deadline. Time mechanics are where multiplayer and persistence constraints bite: a shared world cannot pause.

Information and legibility

Owns nothing in the simulation and decides whether anyone can understand it. Overlays, inspectors, tooltips, dashboards, alerts, advisors, forecasts, fog of war, scouting, and comparative histograms. This family is the counterweight to emergence and the most direct beneficiary of language models that can read state and explain it (§8).

Control and automation

Owns the channel from intent to action: avatars, unit commands, designations, programming, circuits, blueprints, delegation, god powers, and embodied skill. The history of the family is a steady move away from direct control toward specification: from clicking each unit to drawing zones, from wiring logic to writing code, and now toward stating intent in language.

Progression and goals

Owns what the player is working toward: tech trees, milestones, scenarios, career ladders, collections. Simulation games are often sandbox-first, and these mechanics supply direction without scripting outcomes.

Risk and failure

Owns what goes wrong: disasters, wear, permadeath, threat scaling, crime. Failure mechanics are the main source of stakes, and Adaptive threat scaling is the genre's most consequential balancing decision: whether growth makes the game easier or harder.

Narrative and generation

Owns what events happen and how they are told: storytellers, dilemma cards, triggered event chains, requests, chronicles, and runtime-generated content. Emergent narrative needs both a director (what happens) and a chronicle (what it meant). This family is bounded by writing budgets, which is why it is a primary target for generative techniques and the place where their risks (inconsistency, tone, disclosure) are most visible.

World generation

Owns initial conditions: terrain, simulated history, characters. Simulated history generation in Dwarf Fortress simulates centuries of events before play, so the world has a past before the player arrives.

Multiplayer and persistence

Owns other people: persistent worlds, organizations, content sharing, modding, asynchronous regions. These mechanics turn a simulation into a society of players, and they impose the strictest determinism and fairness constraints on everything else.

Conflict and defense

Owns violence as a system: raids and operational military simulation. In most simulation games conflict is a stress test of the economy rather than the point.

Operation and procedure

Owns hands-on work: checklists, tool use on surfaces, diagnosis and repair, rule-checking. The job-sim boom of the 2010s and 2020s made procedure itself pleasurable when tool feedback is good.

5. The structure of the design space

5.1 What games are made of

The average catalog game uses 11.3 mechanics (median 11). The most widely used is Pause and speed control, present in 46 percent of games. The head of the distribution is interface and structure more than simulation: time controls, free placement, and modding support lead, followed by progression structures (tech trees, scenarios) and production chains. Pause controls and modding support are tied to each game's time-model and modding fields (rule E033): every game whose fields imply them tags them, and a few whose fields merely allow them (design-then-test games, content-only modding) tag them too. The table ranks mechanics across families; §4.2 lists them within each family. The figure shows every use at once.

Most-used mechanics across all families. Scores are rubric anchors (1-5).

5.2 How the vocabulary grew