06-five-ceilings 3,943 words about 17 min

Chapter 6. Five ceilings

Your Sim is hungry. You open Buy mode, and the clock stops while you shop. You pick a refrigerator, set it against the kitchen wall, and close the catalog. Time starts again, and the Sim turns from the television, walks across the house, opens the new refrigerator, and makes a meal.

She knew to do that because the refrigerator told her. In The Sims a Sim carries no knowledge of how to use anything. Each object advertises what it offers: a refrigerator advertises Hunger, a bed advertises Energy, a toilet advertises somewhere between 0 and 70 points of Bladder. When a Sim with free will enabled decides what to do next, it gathers the advertisements of the objects it can reach, weights them by its current motives, its personality, and the distance to walk, and picks one. The object then supplies the behavior: the script the Sim runs and the animations it plays [1].

This design, called smart objects, is why The Sims could grow. New objects brought new behaviors with them, largely without changes to the Sims themselves, so expansion packs could extend what Sims did by adding things to buy. The base game shipped with more than 150 objects. It is also why the game's variety had a price that rose in a straight line, and why the design brought its own kinds of bugs: an object whose advertisement is tuned too high becomes a magnet that Sims queue for whatever they need, a broken object that keeps advertising draws Sims into using it again and again, and behavior spread across hundreds of objects is harder to debug than one controller inside the agent. Every behavior a Sim could perform was an object someone had designed, scripted, animated, and tuned. The dataset scores smart objects 4 out of 5 for authoring cost, the level its rubric describes as large content pipelines of animation, writing, and voice.

Authoring cost is one of five limits that have shaped what simulation mechanics can be. The others are the compute a consumer machine can spend on each tick, the amount of hidden state a player can follow, the requirement that the same inputs produce the same world, and how believable a mind built from formulas can be.

Figure: Ceiling scores by family. Mean rubric score of each family's mechanics on four of the five ceilings; determinism is a property of techniques and has its own figure in chapter 11. Blue marks a family at 3 or above on a cost, or below 3.3 on legibility. Families are ordered by combined compute and authoring cost. All scores come from one model rater. Interactive version: https://sim.yohei.me/#fig-constraint-profile

Content made by hand

Some mechanics produce variety from rules alone. A production chain needs a recipe table; a fluid simulation needs a rule for how much water moves between two cells. Others consume content. A dilemma card needs its prompt and options written; an event chain needs every branch written; a social move needs its performance animated. The dataset scores 10 mechanics at 4 or 5 for authoring cost, and the two at 5, the level where content cost is the main limit on the mechanic's depth, are both narrative: triggered event chains and dilemma cards.

Event chains are the backbone of Paradox's grand strategy games. In Crusader Kings III an event fires when the game calls it from a hook such as a birth, a death, or a yearly pulse, if its conditions hold; its options change the state; and it can schedule a follow-up for days or months later, carrying the characters involved along with it [2]. A rival's illness leads to a succession dispute, which leads to an invitation to war. The record for the mechanic lists what goes wrong when the writing runs out: players see the same chains every playthrough once the pool is exhausted, and authored text quality caps depth because each branch must be written by hand.

Dilemma cards fail the same way at a smaller scale. Reigns reduces a kingdom to four meters, church, people, army, and treasury, and a deck of cards that each move some of them; a reign ends when any meter is completely full or completely empty. Over many reigns players meet the same cards again and start answering from memory. Villager requests, the small fetch-and-deliver tasks of Stardew Valley and The Sims 3, are generated from templates to stretch a writing budget, and players learn to read them as chores. Social systems hit the ceiling in their dialogue: a library of social moves that is too small repeats until characters feel scripted, while Prom Week, which went the other way, weighed more than 5,000 influence rules to decide what its characters wanted to do to each other [3].

Open input meets the same ceiling. Façade, the interactive drama from chapter 1, let the player type anything at all to its couple. Behind the text box, the parser mapped each sentence to one of a small set of discourse acts, such as agreeing, criticizing, or flirting, and a drama manager chose the next beat from an authored library; a later analysis counted 30 parameterized speech acts and 27 beats [4]. Any sentence the authors had not anticipated was mapped onto one they had, because every reply had to be written in advance.

Social simulation pays the same bill in interactions. The Sims 3 defined hundreds of social interactions from thousands of production rules [5], and RimWorld rolls chitchat, deep talks, slights, and insults with weights set by opinion and compatibility [6]. The record for social exchanges, which scores 4 for authoring cost, lists what happens when the library is thin or the performance is missing: a small library repeats until characters feel scripted, exchanges that change numbers without a visible performance feel interchangeable, and a player who repeats the best-scoring move maxes out every relationship.

Care that branches runs up the same kind of bill. A creature raised under this mechanic carries a history of how it was treated, and at each stage a table maps that history to its next form. Tamagotchi chooses its adult character from the quality of care, so neglect produces a different creature rather than only a sick one [7]. Every branch is content someone has to design and draw, and the mechanic scores 4 for authoring cost. Its failure modes show what thin branching does: hidden tables push players to walkthroughs, after which they follow a recipe instead of raising anything, and branches that differ only in appearance make the care history feel irrelevant.

Procedural checklists, core to DCS World, score 4 for authoring cost. DCS World's base game is free with two aircraft and two maps; further aircraft and maps are sold as separate modules by Eagle Dynamics and third parties, all running in one client [8].

Machines to take apart are authored part by part. In Car Mechanic Simulator 2021 each car is a tree of modelled parts, more than 4,000 across 72 cars, with a removal order, so a repair means working down to the faulty component before replacing it [9]. PC Building Simulator uses licensed real components with compatibility rules and measures a finished machine with the real 3DMark benchmark, so a build's quality comes from its parts rather than from a score a designer set [10]. Diagnosis and repair scores 4 for authoring cost. One of its failure modes removes the point of the mechanic: when the game highlights faulty parts directly, diagnosis disappears and only the chore of removal remains.

A military simulation's content is its database. Command: Modern Operations draws every platform's sensors and weapons from a database of tens of thousands of real systems, and a Professional Edition sold to defense organizations adds full database editing and Monte Carlo analysis [11]. Military operations score 4 for authoring cost, and the record lists the price of wrong data among the failure modes: database errors or outdated entries undermine credibility with expert players.

The rater who scored the mechanics also judged how open each one is to AI techniques, on a separate scale. The two judgments line up: authoring cost and AI leverage have a Spearman correlation of 0.34 (p <0.001), and the mechanics scored 4 or 5 for authoring cost average 3.5 for AI leverage, against 2.6 for those scored 1 or 2. Since one rater produced both scores, the agreement shows a consistent reading rather than two independent findings. In that reading, the mechanics most limited by what a studio can write and animate are the ones generative models could extend most.

Ticks per second

RimWorld targets sixty ticks a second at normal speed and multiplies that when the player fast-forwards; when the machine cannot keep up, the game runs slower than the speed it shows [12]. The dataset's compute scale puts that experience at its middle: a 3 means a mechanic is a common source of late-game slowdown, a 4 means it is the dominant cost and caps the simulation's scale, and a 5 means it needs specialized hardware or approximation to run at all.

Agent pathfinding produces repathing storms after the player edits the map, a late-game slowdown on large colony and factory maps. Heat moving cell by cell becomes a late-game cost on a large map. A big factory's logistic robots become a compute cost of their own. Rigid-body physics caps the number of bodies in a sandbox. Event scripts that loop over every living character on frequent pulses slow Crusader Kings III late in a game. Fast-forward is where the player meets all of them at once, because it multiplies every other system's cost by the speed.

8 mechanics score 4 or 5. Half are physical: falling-sand materials, fluids and gases, voxel destruction, soft-body deformation. The others are a persistent world shared by many players, simulated history, natural-language interaction, and content generated on demand, the one mechanic at 5. Physics & materials is the family with the highest average compute cost.

The most expensive mechanics in the table are among the least used. Rigid-body physics, which most engines provide, is the most common mechanic in the family, in 13 catalog games; soft-body deformation is in 1, and voxel destruction and falling-sand materials in 2 each. A designer who wants a destructible world in the catalog has few precedents to borrow from.

Teardown and BeamNG.drive show what destruction costs beyond the frame rate. Teardown builds every object from small voxels of wood, plaster, brick, metal, and glass, and chunks cut loose become rigid bodies that fall, roll, and burn [13]. The voxel record lists the other costs: walls and doors cannot stop the player, so Teardown's levels rely on elevation, distance, water, and unbreakable objects, and keeping voxel physics in step over a network was judged infeasible at first, with co-op arriving only in a 2026 update. In BeamNG.drive the node-and-beam skeleton of chapter 3 produces both a vehicle's handling and its damage [14]. Its record adds that the processor cost per vehicle limits how many soft bodies can run at once, and that building a node-and-beam structure by hand for every vehicle is a large authoring cost.

City builders simulate populations as aggregates with a sample of visible people; chapter 5 showed how the original SimCity represented traffic without a single car. SimCity (2013) ran agents that carried every worker and every unit of water between buildings, on cities smaller than earlier games' by design. Per-particle material simulation appears in 2D or in small spaces: Noita simulates every pixel of its world as a material with a state and reaction rules, The Powder Toy pairs its per-particle layer with coarser fields for air pressure and heat, and Oxygen Not Included keeps one element per cell inside a single asteroid. Park and zoo games limit park size, because each guest is an agent with preferences and a path to compute. And Dwarf Fortress moves part of its cost to before play begins: it can simulate thousands of years of a world's history while generating it, which takes a long time and produces very large saves.

A persistent world pays for compute in shared time. EVE Online runs one universe on one server, Tranquility, takes it offline daily for maintenance, and slows the clock in a crowded star system, which it calls time dilation, so that a large fight stays simulated; the battle of B-R5RB ran 21 hours [15]. Second Life simulates each region on its own CPU core with Havok physics, which caps how many avatars and objects one place can hold [16]. The persistent world scores 4 for compute cost.

Screeps makes the compute ceiling a rule of play. Beyond the base limit and bucket of chapter 2, the CPU limit rises with Global Control Level to 300 per tick, and the bucket allows bursts of up to 500 [17]. Scripts are stopped at the limit, so the record notes that players shape their pathfinding and caching around what each costs in CPU.

What a player can follow

The third ceiling is legibility, the subject of chapter 5's measurements. Players have to form a working model of why the simulation did what it did, or they experience its depth as randomness, and on the rubric scores the mechanics that produce the most unscripted variety are rated hardest to read. Its consequence for design is a tax on coupling. Each system a game connects to the others needs an overlay, a tooltip, an inspection panel, or a chronicle to make it readable, and each of those costs design time and screen space.

Same inputs, same world

The fourth ceiling is determinism: the requirement that the same inputs, applied to the same state, produce the same next state every time. Replays work by storing inputs and running the simulation again. Bug reports can be reproduced from a save file. Competitive leaderboards can be checked. And lockstep multiplayer, in which every player's machine runs the full simulation and only inputs cross the network, works only if every machine computes exactly the same world.

Factorio is built around it. Its items move discretely and deterministically, a belt carrying 15, 30, 45, or 60 items a second across two lanes depending on its tier [18]. That makes throughput legible, and it lets every machine in a multiplayer game simulate the identical factory in lockstep.

Opponents that search ahead have the same requirement, and chapter 11 gives the fix: count playouts instead of milliseconds. The Farmer Was Replaced also counts work instead of time, metering its players' programs in simulated ticks (chapter 2), with a successful move costing 200, and its leaderboards run each program from fixed starts for at least two hours and score the average [19, 20].

Determinism has kept randomness that cannot be replayed, and learned models in particular, out of authoritative game state. The dataset sorts every technique that can run during play by what it does to reproducibility. Of the 34 such techniques, 10 break it outright, 16 can be made reproducible by seeding, caching, or logging their outputs, and 8 leave it intact. Infinite Craft, which opens chapter 8, shows the middle category at work: a language model names each new combination once, and the stored answer becomes a rule every later player gets.

Opponents trained by reinforcement learning, such as Gran Turismo Sophy in Gran Turismo 7, and opponents that search ahead, such as Monte Carlo tree search, fall in the same middle category: fix the random seed, or record what the opponent chose, and a race or a match can be replayed. The techniques that break reproducibility outright are mostly the ones that generate behavior and worlds: language-model agents and game masters, realtime voice, and models that generate the next frame of a game from the last. Chapter 11 is about what happens when they meet a game that depends on determinism.

Minds made of formulas

The fifth ceiling is the believability of agents. Many simulated people choose what to do by scoring options. Utility action selection, the dataset's name for the method, enumerates the actions an agent could take, scores each against its needs, the distance, its personality, and the context, and picks the best or samples among the best; Theme Park used it in 1994, and The Sims chose each behavior by how much it would raise a Sim's happiness. The Sims 3 added a deliberate touch of randomness: its Sims sample among good options rather than always taking the best one, and sample more widely when they are doing poorly. The method is cheap and legible. Two actions with close scores alternate, so agents start and abandon tasks. One consideration dominates the product, and agents ignore everything else. And since each score is the product of many curves, players see randomness where there is arithmetic.

Designers have worked within the ceiling in a few ways. One is many shallow agents. A RollerCoaster Tycoon guest has money, a preferred intensity, a nausea tolerance, and a few needs, and the park is judged by thousands of them at once [21]. The other direction is a few agents with thick state: the itemized memories of RimWorld and Dwarf Fortress, or bodies that remember every wound. A third direction tried to grow minds instead of writing them. Each Norn in Creatures (1996) had a neural-network brain, a simulated biochemistry that set its drives, and a digital genome that built both; drive reduction reinforced whatever the Norn had just done, so the player's tickles and slaps taught it only through the same chemistry. Because the genome built both the brain and the body, mutations produced Norns that could not see or never died, and players traded them online. The approach made creatures that surprised their owners and stayed hard to direct, and the catalog's record of it has few successors: creature learning appears in 4 games.

A fourth direction kept the authored social world and gave agents a way to look ahead. In Versu, built by Richard Evans and Emily Short and released by Linden Lab in 2013, an agent tests each candidate action by applying it to the world state, scoring the result against its desires, more than 300 in all, and undoing it; the authors contrast this with the estimated consequences The Sims relies on [4]. Situations such as a meal or a conversation are written as social practices that offer actions to the agents taking part without controlling them, and a stress tester that ran hundreds of simulated instances supported more than 1,000 parameterized actions. Linden Lab cancelled the project in early 2014.

Tomodachi Life takes its characters from the player. Its Miis have the looks, names, voices, and personalities the player gave them, and most of its comedy comes from the pairings and quarrels they start on their own [22]. The player cannot direct them and instead answers their requests, and key changes in a relationship pass through the player's consent: a confession happens only after the player approves it.

Rich minds, agents that converse, plan, and remember in open language, sat outside all four directions until recently.

Generative Agents, a 2023 research sandbox, put 25 language-model agents in a town modeled on The Sims. Each kept a memory stream of observations in text, retrieved memories by recency, importance, and relevance, wrote reflections two or three times a game day, and planned its days in chunks from hours down to minutes. Told that one agent wanted to host a Valentine's Day party, the town spread the news by conversation: over two game days 13 of the 25 agents heard about it and five attended [23]. The same approach raises believability and spends two other ceilings to do it: each agent's choices cost inference at runtime, and none of them can be reproduced from the inputs alone. inZOI, a commercial life sim in early access since March 2025, puts a smaller version of the idea on the player's own machine: a small language model running locally lets typed prompts steer a character's thoughts and actions.

Lifting one ceiling spends another

Smart objects let The Sims add behaviors without touching its agents, and every behavior became an object to author. Individual vehicles let Cities: Skylines show a jam forming at a particular intersection, at a cost in compute and in a lane-choice model that players could not read. Agents that carried goods in SimCity (2013) made the city's flows visible, on maps smaller by design than earlier games'. Generative Agents bought conversation and memory with inference and with reproducibility. Infinite Craft bought open-ended crafting with a language model and kept determinism by caching every answer.

The dataset's scores give a rough map of where each mechanic sits on each ceiling, and the techniques in chapters 7 and 8 change the prices.

The ceilings move at different speeds

Authoring cost is giving way fastest. Asset generation, learned motion, and language characters have all shipped in commercial games, and the techniques that write content offline, before a player sees it, keep determinism intact or can be made reproducible by storing what they made. Compute is giving way mostly in rendering. Neural upscaling and frame generation are the one AI technique in the dataset that is a norm in shipped games, and they never touch simulation state, which is part of why they could spread. Learned substitutes for the simulation itself, for physics, fluids, and weather, remain research or prototypes in games. Agent believability is giving way at a cost in determinism and legibility, as Generative Agents shows. And for legibility, explanation is the rarest effect any technique in the dataset has on a mechanic, as chapter 5 showed.

Exercise: find the ceiling

Pick a mechanic from a game you know that you wish went deeper: villagers with more to say, a city with more people, a physics model with more pieces, a rival who remembers what you did.

1. Name the ceiling that stops it: authoring cost, compute, legibility, determinism, or believability. Some mechanics hit two.
2. Find the mechanic in the dataset and check its rubric scores. Do they agree with your diagnosis?
3. Describe the cheapest change that would lift the ceiling a little. Write down what it spends: another ceiling, money per player hour, or the designer's control over what happens.
4. Name a game that made the opposite trade, keeping the mechanic shallow to spend its budget elsewhere.

Chapter 7 follows the techniques that press on these five ceilings into shipped games. It begins in 2019 with AI Dungeon, a text adventure with no rules engine underneath.

Sources

[1] Kenneth D. Forbus, Will Wright (2001-05-31). Some notes on programming objects in The Sims. Qualitative Reasoning Group, Northwestern University (course notes) https://www.qrg.northwestern.edu/papers/Files/Programming_Objects_in_The_Sims.pdf
[2] CK3 Wiki contributors (2026). Event modding (Crusader Kings III). Crusader Kings III Wiki (paradoxwikis.com) https://ck3.paradoxwikis.com/Event_modding
[3] Joshua McCoy, Mike Treanor, Ben Samuel, Aaron A. Reed, Michael Mateas, Noah Wardrip-Fruin (2014-06). Social Story Worlds With Comme il Faut. IEEE Transactions on Computational Intelligence and AI in Games 6(2) https://cs.uky.edu/~sgware/reading/papers/mccoy2014cif.pdf
[4] Richard Evans, Emily Short (2014). Versu: A Simulationist Storytelling System. IEEE Transactions on Computational Intelligence and AI in Games 6(2) https://www.cs.uky.edu/~sgware/reading/papers/evans2014versu.pdf
[5] Richard Evans (2010). Modeling Individual Personalities in The Sims 3. Game Developers Conference 2010, Programming track https://www.gdcvault.com/play/1012450/Modeling-Individual-Personalities-in-The
[6] RimWorld Wiki contributors (2026). Social. RimWorld Wiki https://rimworldwiki.com/wiki/Social
[7] Wikipedia contributors (2026). Tamagotchi. Wikipedia https://en.wikipedia.org/wiki/Tamagotchi
[8] Wikipedia contributors (2026). DCS World. Wikipedia https://en.wikipedia.org/wiki/DCS_World
[9] (2026). Car Mechanic Simulator 2021 on Steam. Steam https://store.steampowered.com/app/1190000/Car_Mechanic_Simulator_2021/
[10] Wikipedia contributors (2026). PC Building Simulator. Wikipedia https://en.wikipedia.org/wiki/PC_Building_Simulator
[11] Wikipedia contributors (2026). Command: Modern Operations. Wikipedia https://en.wikipedia.org/wiki/Command:Modern_Operations
[12] RimWorld Wiki contributors (2026). Time. RimWorld Wiki https://rimworldwiki.com/wiki/Time
[13] Wikipedia contributors (2026). Teardown (video game). Wikipedia https://en.wikipedia.org/wiki/Teardown(video_game)
[14] Wikipedia contributors (2026). BeamNG.drive. Wikipedia https://en.wikipedia.org/wiki/BeamNG.drive
[15] Wikipedia contributors (2026). Eve Online. Wikipedia https://en.wikipedia.org/wiki/Eve_Online
[16] Wikipedia contributors (2026). Second Life. Wikipedia https://en.wikipedia.org/wiki/Second_Life
[17] (2026). How does CPU limit work. Screeps documentation https://docs.screeps.com/cpu-limit.html
[18] Factorio Wiki contributors (2026). Transport belts/Physics. Official Factorio Wiki https://wiki.factorio.com/Transport_belts/Physics
[19] The Farmer Was Replaced Wiki contributors (2026). Operation Costs. The Farmer Was Replaced Wiki https://thefarmerwasreplaced.wiki.gg/wiki/Operation_Costs
[20] The Farmer Was Replaced Wiki contributors (2026). Leaderboard. The Farmer Was Replaced Wiki https://thefarmerwasreplaced.wiki.gg/wiki/Leaderboard
[21] OpenRCT2 contributors (2026). OpenRCT2 source: Guest.cpp (guest ride choice and thoughts). GitHub (OpenRCT2/OpenRCT2) https://github.com/OpenRCT2/OpenRCT2/blob/develop/src/openrct2/entity/Guest.cpp
[22] Wikipedia contributors (2026). Tomodachi Life. Wikipedia https://en.wikipedia.org/wiki/Tomodachi_Life
[23] Joon Sung Park, Joseph C. O'Brien, Carrie J. Cai, Meredith Ringel Morris, Percy Liang, Michael S. Bernstein (2023-04-07). Generative Agents: Interactive Simulacra of Human Behavior. UIST '23 (ACM Symposium on User Interface Software and Technology) https://arxiv.org/abs/2304.03442

Dataset dfde7895aa6d.

mechanics in this chapter

24 linked or named

Agent pathfindingCreature learningDilemma cardsFalling-sand materialsFlight aerodynamicsFluid and gas simulationHeat transferMaterial interactionsMilitary operationsNatural-language interactionOrbital mechanicsProcedural checklistsProjectile ballisticsRigid-body physicsSmart objectsSocial exchangesSoft-body deformationTerrain deformationTriggered event chainsUtility action selectionVehicle dynamicsVillager requestsVoxel destructionWeather