People and histories

How a generated place can come with people who have a past: names in a language of their own, cultures read from the land, a history that explains the ruins and shrines, and days that run on a clock. Tyrell already runs the minds: needs, routines, memories, households, lineage and gossip. What it lacks is a generator to write those people before minute 0, and that is this page. The finding is that every shipped system splits into the same two halves. History is either simulated, then summarized (Dwarf Fortress, Talk of the Town) or generated as events, then explained (Caves of Qud), and for a quest-sized world the second is cheaper and reads better. Either way, what matters is that every event leaves something in the place and in someone’s memory. Names work the same way: a language is a few phoneme sets, a syllable shape and a spelling (Martin O’Leary), and if each morpheme keeps its meaning, a place can be named from what is there and glossed back. Routines are data in every game that has them, from Ultima VII’s eight three-hour slots to Stardew Valley’s keyed variants. Tyrell’s routine format already is that data, so a generator only has to write it. A 300-year history for a small island takes about 70 ms of plain JS, and a village’s day is a pure function of the clock.

Proposed: none of this is in Tycho yet. The demos run here in the page, written in plain JS with the study’s own kit (demos/kit.js), not Tycho; each says how it would run in Tycho: per frame (the GPU stand-in’s calls), pre-bake (painted once offscreen), or shader (per-pixel work every frame). What would move into Tycho, and when, is in the roadmap.

Text marked (inference) is reasoning from the sources, not a sourced fact. “(search summary)” marks a claim known only from a search-result snippet, where the source itself could not be fetched. The full research is in the note on people, names and histories.

• What Tyrell already does
• Names from a naming language
• Place names from the land
• Cultures from the land
• History: simulate or explain
• Lineage and legacy
• Daily routines
• Off screen
• A quest-sized past
• What it means for Tycho and Tyrell
• Sources

What Tyrell already does

Tyrell decides what each person does, remembers what happened to them and “keeps them living when nobody is watching” (Tyrell). For a generator, the useful facts are what a person is in Tyrell, and which inputs can give them a past.

Tyrell knows “nothing about tiles, side views, sprites or maps” (Tyrell), and it never invents a name, a language, a culture or a past. That is the gap. Everything below would be a generator that runs once, from a seed, and writes plain JSON that Tyrell’s existing inputs accept (inference).

Names from a naming language

Martin O’Leary’s naming-language generator, made for his Uncharted Atlas map bot, is the reference design, and it is small: one file of JavaScript (mewo2/naming-language). A language is a handful of choices:

Two details make the output sound like a language rather than noise. First, frequency is skewed: each language shuffles its phonemes once and picks with list[floor(random()² × length)], so every language has its own common and rare sounds. Rosenfelder’s Gen does the same with a 30% chance of stopping at each phoneme, which gives the first four about 30, 21, 15 and 10% (Gen help). Second, morphemes are kept by meaning. A keyed concept such as “city” reuses an existing morpheme with probability n/(n + 1), so it settles on one or two stems, and the genitive “of” and the definite “the” are one morpheme each per language. A name is one word or two, joined by a space, a hyphen or the genitive, kept between 5 and 12 letters, and rejected if it contains another name. Rosenfelder’s advice is the same in prose: “Start with a distinctive syllable pattern”, and “It’s better to derive long words by combining shorter words, or adding suffixes” (Language Construction Kit).

The alternatives trade control for flavor:

The demo combines the first and third rows. It keeps O’Leary’s sets, structures, restrictions and spellings, but gives every morpheme a meaning (“river”, “ford”, “oak”, “wolf”). It makes them all once, in a fixed order, so a language is the same whatever asks first. Then any name can be glossed back. Dwarf Fortress shows the payoff: a dwarf’s surname is two English words rendered in dwarvish, “Urist Borushdumat (Raked-Roughness)” (search summary, DF wiki), and 60-odd kinds of thing can be named, roads, bridges and tombs among them (DFHack’s df-structures). Personal names in the demo are two meaningful elements, as Old English names were (Ælfred, “elf counsel”) (inference from common knowledge; not separately sourced). Family names follow one rule per language: by father or mother, by home place, by trade, or by a house name. One rule per language is what makes a household read as one family and a culture as one culture (inference).

Place names from the land

Real place names are a generic for the kind of place plus a specific that tells it apart (inference from the lists below). The English generics are a small table (Wikipedia; the Key to English Place-Names gives each name’s elements and their languages):

Three things carry over to a generator (inference):

• The generic comes from the map. A town on a river cell with a road across it is a ford; a farm by the sea is a shore farm. In the demo every site is classified by its cell and neighbors, the same way the settlements page scores village sites.
• The specific comes from the neighborhood or the history: a tree from the nearby wood, salt from the sea, a color, an animal, or a founder (“Beorma’s people’s homestead” is how Birmingham is usually glossed). The history demo names about one town in three after its founder.
• The table can be swapped per culture. English is head-last (-ford); Welsh and Gaelic names are head-first (Aber-, Llan-, Inver-). Old Norse and Old English generics in one county record who settled where. A map where two cultures meet should show two sets of generics, and a town that changed hands can keep its old name.

O’Leary’s own map generator calls makeName(lang, 'city') and makeName(lang, 'region') (mewo2/terrain): his keys repeat a stem without saying what it means. Glossing is the one addition. It lets the game say what a name means when someone asks, and lets a quest turn on a name (“Oakford has no oaks”) (inference).

Cultures from the land

Generators that make cultures start from where the culture begins. Azgaar’s Fantasy Map Generator reads a culture’s type from its center cell, first match wins: Nomadic (low desert or grassland), Highland, Lake, Naval (a harbor), River (flux over 100), Hunting (inland forest, taiga, tundra, wetland), or Generic. The type sets how far it spreads (expansionism 0.7 for Hunting to 1.5 for Naval and Nomadic) and what each cell costs it: River cultures pay 0 along rivers and 100 away from them; Highland cultures pay 3,000 for lowland; Naval cultures cross water at a twenty-fifth of a Nomadic culture’s cost (cultures-generator.ts). Each culture points at one name base, and folk religions come from their culture.

What a culture values is a short vector in every system that models it:

Mark R. Johnson’s Ultima Ratio Regum goes furthest, and he names the danger: most cultural generation is “tremendously simplistic… nothing more than surface detail”. His fix is to make every trait refer to something else in the world. A nation’s dialect draws on its capital’s climate, terrain, plants, animals, history and religion: a nomad’s curse is “may you be buried beneath the sand where none shall find you”. Sentence complexity is gameplay, because talkative cultures give more clues (Johnson 2016). His other habits keep the output from becoming a set of stock types. Outcomes are skewed: “three outcomes which each happen 30% of the time” beside “50 incredibly obscure ones which only happen 0.2% of the time”. A hidden check asks whether something “is too much like something else” and rerolls it. And “every concept should in some way link to every other concept”: if “ravens are thought to be lucky”, a raven on a shield earns approval (Game Developer).

For Elsewhere, whose quests each bring a small world with “its own look, people and way of playing”, a culture can be a short record (inference):

The demo’s cultures are built this way. The land adds weight to some values (seafarers lean toward trade), but every value can come up anywhere. Taboos and building styles are drawn separately, so two seafaring cultures don’t come out identical. That is a small version of Johnson’s skew and interlinking, and it is the honest limit of the demo: four fields don’t make a culture, they make a culture’s label (inference).

History: simulate or explain

Simulate, then summarize. Dwarf Fortress builds the land, then “the simulation of civilization begins. Through this process settlements are raised, trade routes are formed, and wars are waged. At a designated point in time, the simulation stops and the player begins” (Tarn Adams, Game AI Pro 2). History length runs from 5 to 1,050 years, and more civilizations and sites lengthen generation “by a huge amount” (search summary, DF wiki). The record is kept as data: historical figures with birth and death years, events with a year, a type and the ids they touch, and collections (a war, a battle, a journey) that nest events. DFHack counts 134 event types (df-structures). Tarn’s rules for the model are short: “Don’t Overcomplicate”, and “Operate at the level of what the player sees or one layer below”. He also asks of each feature “whether or not a story thread can pass through” it (Paprika).

Talk of the Town simulates a town “from its founding in 1839, when a handful of families converge on an empty townscape to establish farms, through the death of the central character in the summer of 1979”, ending with 300 to 500 people (Ryan and Mateas). Its cost tricks are the lesson:

• It simulates only ten half-days a year, about 1.4% of them (basic_config.py).
• It skips knowledge for 140 years. A week before the end it implants what each person knows of family, friends, neighbors and coworkers, then simulates one week in full.
• Even so, worldgen takes “a few minutes”.

Ryan’s dissertation then argues that the raw run “will almost always lack story structure”, so it must be sifted and curated, the way a nonfiction writer would (Emily Short’s review).

Generate the events, then explain them. Caves of Qud’s sultan histories start from “a distinction to be made between history as a process and history as an artifact”, and aim “to generate history in order to simulate historical accounts” (Grinblat and Bucklew 2017). A sultan is a small state (name, birth year, region, a domain such as ice or glass, allies, a profession). About twelve events are drawn from a pool, and “there’s no logic behind the choice of events… though their gospels often profess causes.” Each event’s text, “Acting against #injustice#, #sultanName# led an army to the gates of #location#”, fills its cause from the state: if the sultan has allied frogs, the injustice is “the persecution of frogs”. If no state fits, the event invents it and adds the frogs: “the effect causes the cause.” Coherence comes from the few shared properties: “The sultan’s shared state acts as a glue.” Every event leaves things in the world. Visited places become historic sites, named items lost there lie there as real loot, and shrines and murals each show one gospel. Tomb inscriptions retell the same events “more laudatory”, a second, biased account (Qud wiki). A village gets only “two generated events”, each with a 1 in 20 chance of having emptied it, leaving “sometimes explicit clues about what happened” (Qud wiki).

Between the two sit story sifting and constraint solving. Kreminski’s Felt runs Datalog patterns over the event log to find “events that constitute a compelling story”, and reads Qud as sifting “to retroactively justify each action” (Felt). Winnow sifts while events are still playing out, so a game can foreshadow (Winnow). Horswill’s CatSAT solves for a consistent character instead (CatSAT). And Nothing Beside Remains runs “an abstract simulation of a village population” until it fails (food, ecosystem, hostile animals), then prints the end state into the map: a hot, dry end means shrunken water (Smith Nicholls and Cook 2023).

The demo is the simulate side made cheap. It ticks yearly, never daily, keeps no agents beyond towns and rulers, and stores every event with its year and the ids of what it touched. Three hundred years on a 96×60 island with three or four cultures take 60 to 75 ms in Chrome on a laptop. Generation follows these rules:

• Towns grow toward what their site can feed.
• Full towns found daughters on the best free site nearby. The road between them is found by cost, so roads bend around hills and cross rivers at fords, and later roads reuse them.
• Neighbors drift toward trade or quarrel, by their values and by land hunger.
• Wars produce battles, captured towns and ruins. A fallen hero gets a cairn.
• Floods produce shrines upstream, and visions produce shrines on hills.

Because each event names what it touched, tapping a shrine on the map answers why it is there.

For a quest the explain side is usually the better fit (inference). The player arrives at the end of history and never stops it partway, a quest needs three to five events rather than three hundred years, and each event can be chosen because it leaves something worth finding. The demo’s real lesson is its data shape: events as { year, text, refs }, with every object on the map holding the ids of the events that made it. That shape serves either method.

Lineage and legacy

A past is also a family tree. Crusader Kings III authors its history and simulates forward from it. Character files carry dynasty, culture, father and mother by id, traits, and dated birth and death blocks (CK3 wiki). Congenital traits pass on by table: 25% with one parent showing it, 80% with both, and 10% from two carriers (CK3 wiki). Tyrell’s Cradle and Delos already do the forward part: ages and stages from born, couples, births whose genes follow dominance and whose traits blend both parents (Delos). So a generator needs only to write the founders and the dead: a few generations as PersonDefs with negative born, and the departed as memories of 'loss' (inference).

Wildermyth shows how little text a character needs. Each hero starts with three history lines (Origin, Anecdote, Motivation), “A line of player-facing text that tells a story about the entity it’s pinned to”. The lines set stats and seed quirks, and more are added “in recognition of game-occurrences” (Wildermyth wiki). A retired hero becomes a legacy hero who keeps “items, transformations, history, and abilities” (Legacy). That is the right size for an Elsewhere quest-giver: one or two lines of past, each tied to something the player can see (inference).

Daily routines

Routines are data in every game that has them, and the formats converge:

Tyrell’s format is Ultima’s entries with Stardew’s variants, run over The Sims’ needs, which leaves only the writing of routines to a generator (inference). The demo writes them from a few parts:

• The job: where and when work happens. The smith works the forge 7 to 17, the baker from 4:30, the innkeeper until 23:30. The trader walks the east road except on market days; the herder takes the flock to the west hills.
• The stage of life: children play by the well, and elders sit there in the morning and nap after lunch.
• The person’s own habits: a lark or owl score sets waking (as Animal Crossing’s personalities do), devoutness adds a dawn prayer, and sociability chooses the tavern, the well or home in the evening.
• The village’s values: with kin, supper is at 18:00 sharp for everyone; with faith, more people pray; with merriment, more go to the tavern.
• Two weekday variants: market days (1 and 4) bring traders and buyers to the square, and the rest day (6) brings everyone to the shrine at 9.

Each day adds a few minutes of seeded slack per entry, so nobody runs like a clock, except the shared supper and the rest-day prayer, which are social appointments. Bethesda’s history is the warning against going further. Todd Howard promised NPCs that “can steal it, they can buy it, they can grow it”; Emil Pagliarulo found a quest’s skooma dealer already dead, because addicts “were killing the merchant to get it!”, and the team “had to consciously tone down” behavior (GamesBeat). Shipped, “NPCs cannot buy food” (Paavo). A routine is a promise to the player about where someone will be; needs should bend it, not break it. Tyrell’s ranks already do that.

Off screen

Every shipped system cheats the same way off screen:

Tyrell’s answer is exact rather than approximate. “Off screen nothing ticks”: an event queue jumps between action ends, so “a day for a hundred people is a few thousand events, not a few million frames”. Near and far follow one timetable, so a watched world hashes the same as an unwatched one (Concepts). Strangers go one step further and cost nothing until promoted (Places). The schedules demo is the strangers’ trick applied to routine-only people: their position at any minute is a pure function of the seed and the clock, with nothing stepped. For history, the same idea runs one level up (inference):

• Years before the quest are a coarse simulation or a written account, done once at load.
• Days before the quest are Tyrell’s advance over a day or two, so habits, regard and gossip settle. A night for Rocinante takes “tens of milliseconds” (Time).
• The quest itself runs live.

A quest-sized past

Elsewhere’s quests are “five to fifteen minutes”, each with “two to six residents” and one place (Elsewhere’s design brief, its DESIGN.md). The worlds page already proposes the loop: a few people, three to five events that link people to places, a layout where each event leaves something, and the events fed to the minds as memories. Put together with this page’s sources, the generator would run in seven steps, all from one named stream such as ctx.rng('people') (inference):

1. Read the site. From the land the place already has (geography, settlements): water, slope, woods, the road and its fords.
2. Make the culture. A way of life from the site, two values, a taboo, a building style, a language and its naming rule.

3. Write the past backward from what the place shows. Choose three to five events from a deck whose cards each leave a mark, Qud-style. Fill each card’s cause from the state, or invent it:

   Card	Leaves	Becomes
   Flood	A shrine upstream	A fact, a memory for everyone old enough
   Fire	A ruin, a rebuilt forge on old stones	Who owns the forge now
   Quarrel over a boundary	A field split by a wall	Two households with a grudge (Fellowship sentiment)
   Marriage from elsewhere	A newcomer with a different name	A name from another language, a sentiment
   Death	A grave	Grief, an inheritance (Cradle)
   A road built round a hill	The bend in the road	A name (“Long Hill”), a memory

4. Name everything from the site and the events, in the culture’s language, glossed.
5. Write the people. Households with born, home, a job and a routine from the parts above, values from the culture plus a personal draw, and the events as reverie.remember memories for those who lived them, each with about and place.
6. Settle. world.advance a day or two so regard, habits and gossip start from somewhere.
7. Check. In Node over a thousand seeds: every person has a reachable home and workplace, no two names collide, every event has a mark the player can reach, and generation stays under its budget. This is the generate-and-test the worlds page recommends.

Why the shrine is there, why the road bends, who owns which field. Each of these is an event whose mark is on the map and whose memory is in someone’s head. The player can find it in either place, and Voice can bring it up when they ask (inference). Dwarf Fortress’s test for a feature fits here: whether “a story thread can pass through” it.

What it means for Tycho and Tyrell

Nothing here needs new engine machinery for minds, because Tyrell already runs everything the past is fed into. What’s missing is a seeded writer of people, small enough to run at a quest’s load. A proposed split, all (inference):

Tyrell itself might gain two small options (inference):

• A routine variant keyed by a fact the chronicle sets (a feast on a flood’s anniversary). when: ['fact:...'] already covers it.
• A memory kind for “heard of it” as against “was there”, so the young know the flood only as a story. Reverie’s beliefs with sources come close.

Elsewhere’s rule still holds. Nobody in any world knows what a main quest is, and no generated past may be about the Hollow, the Lanterns or the Spire. The event deck simply has no such cards.

Sources

• Tyrell: README, Concepts, Places, Time, Deciding, API, and the modules Reverie, Fellowship, Cradle, Delos
• Martin O’Leary: naming-language and terrain
• Mark Rosenfelder: Language Construction Kit, Gen and its help
• Colin Gorrie, conlang word generators; Lingua::Awkwords; Vulgar; fantasyname; Frozen Fractal, constructing languages
• RogueBasin, Markov names with Katz back-off
• Azgaar’s Fantasy Map Generator (names, cultures and religions generators)
• Wikipedia, generic forms in place names in the UK and Ireland; Key to English Place-Names
• Dwarf Fortress: Tarn Adams, Game AI Pro 2 ch. 41; Paprika interview; DFHack df-structures; DF wiki: World gen, Language, Urist
• Grinblat and Bucklew, Generation of Mythic Biographies in Caves of Qud (FDG 2017); Qud wiki: Sultan history, Village
• Ryan and Mateas, Talk of the Town, Game AI Pro 3 ch. 37 and its config; Ryan, Curating Simulated Storyworlds, via Emily Short’s review
• Kreminski: Felt, Winnow; Horswill, CatSAT; Smith Nicholls and Cook, Nothing Beside Remains
• Mark R. Johnson: Procedural Generation of Linguistics, Dialects, Naming Conventions and Spoken Sentences (2016); Game Developer, ten years of Ultima Ratio Regum
• Stellaris wiki, Ethics; CK3 wiki: Characters modding, Trait inheritance; Wildermyth wiki: History, Legacy
• Ultima Codex, Ultima VII SCHEDULE.DAT; Paavo, What was Radiant AI, anyway?; cs.uesp.net, Package; GamesBeat, dimming the Radiant AI
• Forbus and Wright, programming objects in The Sims; RimWorld wiki, Schedule; Nookipedia, Personality
• Shadows of Doubt: DevBlog 8, DevBlog 15; Assassin’s Creed Unity crowds, GDC 2015

The full research is in the note on people, names and histories.