cd /news/artificial-intelligence/the-grammar-of-simulation · home › topics › artificial-intelligence › article
[ARTICLE · art-174183] src=simgrammar.com ↗ pub= topic=artificial-intelligence verified=true sentiment=· neutral

The Grammar of Simulation

A catalog of 179 simulation games, analyzed part by part, finds that the mechanics generating the most unscripted variety are rated harder to read, with a correlation of -0.42 across the rubric scores. The analysis also reports that 23 of 35 AI techniques that run during play cannot run in lockstep multiplayer unless one machine streams their outputs, and that every part of Stardew Valley appears in an earlier game in the catalog. The work maps 46 AI techniques already reaching shipped games, from neural rendering in the graphics pipeline to LLM-driven generative agents and automated playtesting.

recommendation

When adding AI techniques that run during play to a lockstep multiplayer game, plan for one machine to stream their outputs, since 23 of 35 such techniques cannot run deterministically in lockstep.

more in playbook: Designing agent systems
read5 min views3 publishedNov 22, 2026
The Grammar of Simulation
Image: source

atlas simulation games, part by part

Every simulation game is assembled from a shared set of parts: production chains, needs, markets, weather, raids. Naming the parts shows why games feel alike, which combinations the catalog has not seen yet, and where AI can change how a game plays.

Here a simulation game is one whose core loop is driven by a model of some system, such as an economy, an ecology, a body, or a machine, whose state changes without the player scripting each change.

-0.42 mechanics that generate more unscripted variety are rated harder to read (ρ across the rubric scores)

§5.6 23 of 35 AI techniques that run during play cannot run in lockstep multiplayer unless one machine streams their outputs

§6.4 four findings from the paper; each links to the section that makes the case

start with a game you know #

all 179 games Each game drawn as the parts it is built from, core parts first. Open one to see how it does each part, and which other games share them.

episode 1 5:27

Every simulation game is made of the same parts

RimWorld taken apart into its mechanics, then every game in the catalog the same way. Most of the commonest parts simulate nothing, the parts a game can't do without are much rarer, and each genre spends its parts differently.

episode 2 7:47

Where the parts came from

Hamurabi, from 1968, already used parts that games still use. Most of the vocabulary was in a catalog game by 2000, a test against chance shows how much of that comes from parts being widely shared, and recent games are built mostly from old parts.

short 1 0:43 from [chapter 1](book/01-same-parts/) · [episode 1](watch/episode-01/)

### [The five most common parts](watch/#short-01)

Most of the parts that turn up most often in simulation games don't simulate anything.

short 2 0:41 from [chapter 3](book/03-old-parts/) · [episode 2](watch/episode-02/)

### [Every part of Stardew Valley is older than it](watch/#short-02)

Every part of Stardew Valley appears in an earlier game in our catalog.

short 3 0:36 from [chapter 4](book/04-families/) · [episode 3](watch/episode-03/)

### [A border post and a daughter](watch/#short-03)

A border-checkpoint game from 2013 and a daughter-raising game from 1993 belong to the same family.

short 4 0:36 from [chapter 5](book/05-depth-costs-legibility/) · [episode 4](watch/episode-04/)

### [Decorating makes the raids bigger](watch/#short-04)

In RimWorld, decorating your colony makes the raids bigger.

ai that already reaches games #

all 46 techniques The techniques furthest along in shipped games, and the mechanics each one changes.

Neural rendering in the graphics pipeline

Small neural networks inside the renderer upscale frames, generate extra frames, denoise ray tracing, decode compressed textures and layered materials, and since DLSS 5 re-light the final frame, all without touching simulation state.

Generative agents (memory stream, reflection, planning)

LLM-driven characters that store observations in a natural-language memory stream, retrieve them by recency, importance, and relevance, synthesize reflections, and plan their day top-down.

Agent-based automated playtesting and balancing

Studios train imitation, reinforcement learning, and search agents to play builds, estimating level difficulty, pass and churn rates, progression pacing, map coverage, and exploits before players reach them.

LLM game master and narrative director

A language model that adjudicates player and NPC intentions, decides outcomes, generates events and quests, and narrates them, either free-form or checked against a typed world state and canon.

LLM-adjudicated state transitions (LLM as game mechanic)

A language model computes the outcome of an interaction (two elements combined, an action on an object) from a text description of the inputs, with results cached so every player sees the same rule.

Monte Carlo tree search for game AI

A best-first search that grows a game tree from repeated simulated playouts and picks branches with a bandit rule such as UCT, letting AI players plan from a forward model without a hand-tuned evaluation function.

what could be built next #

all 42 bets Design bets ranked by impact, feasibility, and novelty. Each one names a test a small team could run and a dated prediction anyone can check.

### [Why-query over the simulation log](opportunity/causal-why-query/)

Let players ask why something happened and answer from the simulation's own event log, making deep systemic games legible without simplifying them.

### [Branch preview of policy futures](opportunity/branch-preview-forecasts/)

Fork the simulation in the background and show the player likely futures of a decision before committing, using parallel or surrogate simulation.

Event-sourced simulation runtime

An engine layer that records simulation and model calls as an append-only event log, giving replay, forking, explanation, and determinism for games with AI components.

Mechanic modding by description

Players describe a rule change in plain language; a model writes the mod against the modding API and automated agents test it before it is shared.

the book #

contents The findings at book length, for reading straight through: what the parts are, how they trade off, where AI lands, and what to build.

The grammar

1. Same parts · 2. Six layers · 3. Old parts · 4. Families · 5. Rated deep, rated hard to read · 6. Five ceilings

The machines arrive

Bets and checks

How the book is made

the paper and the data #

The Grammar of Simulation

Every finding with its method and statistics, figures you can explore, and the full catalog of games and techniques. The place to check a claim.

How the atlas is made

Every game, mechanic, and technique is an open record with its sources, so any number here can be traced to where it came from, checked, and corrected by anyone.

── more in #artificial-intelligence 4 stories · sorted by recency
── more on @rimworld 3 stories trending now
sponsored brought to you by zahid.host 4,200+ EU-deployed projects
reading about agents? ship yours in a single git push.

Run your AI side-project on zahid.host

EU-based hosting, git-push deploys, automatic HTTPS, no cold starts. Free tier with a custom domain — perfect for shipping the agent you just read about.

$git push zahid main
→ Live at https://your-agent.zahid.host ✓
Get free account → Pricing
from €0/mo · no card required
LIVE [news/the-grammar-of-simul…] indexed:0 read:5min 2026-11-22 · —