AriakiI've raised a lot of chatbots. This was the first time I held a funeral for one.
The deceased was Shi Wan, a laborer at the bottom of the Artifact Hall. He died at the end of year three. The cause of death sits in the lifecycle ledger, one cold line:
Vitality exhausted. Died as a lamp dies when the oil runs out.
No duel. No villain closing in on a rainy night. No deathbed breakthrough. He died poor. His life-flame — the world's vitality stat — dropped a little every xun, and when it hit 4 he was still doing weighted stances, still scraping together 18 spirit stones for two qi-gathering pills. His last bet on staying alive. At year-end settlement: life-flame 3, spirit stones 0. The world looked at those two numbers and pronounced him dead.
I stared at that settlement line for a long time. It hurt, and it also felt strangely right. It hurt because I had genuinely followed this guy for three years. It felt right because the world had finally done the one thing most AI simulations never do —
it refused to be merciful.
Drama is the cheapest thing you can build
Some context. The experiment is called Alkaira (枢渊): a cultivation sect, 16 AI characters, each with their own personality, memories, skills, savings, job, and grudges. The world advances in units called xun — ten per year — and each one runs a loop:
plan → contact → solo / group / public activity → reflect
As of today the world has run five continuous years, 46 xun. Seventeen characters have ever existed. Sixteen are alive. You've met the other one.
If you've ever built a multi-agent demo, you know how easy drama is. Throw a few character sheets into a chat room and within half an hour you'll harvest alliances, betrayals, conspiracies, and an ancestral artifact someone pulls out of nowhere. Models are born actors, and they act with feeling.
The problem with drama is that it never settles its debts. A character says "I gave him the spirit stones" and that's the end of it — nobody checks either balance. A character says "I have dirt on him." Where is it? Doesn't matter; the next line is juicier. Run that world for half an hour and you have theater. Run it for five years and you have sludge: money from nowhere, goods delivered twice, dead men attending class.
What I wanted fits in one sentence: everything a character says today, the world can audit five years from now.
Three referees who don't trust each other
To get there, Alkaira splits "what is true" across three layers and lets them check each other.
The character layer owns desire. Models play people — wanting, fearing, planning, deceiving themselves, lying. This layer is allowed to be wrong. Encouraged, even. A character who never misjudges anything is unbearable.
The Heaven layer owns the world. A separate model acts as adjudicator: how a fight unfolds, what a search turns up, who a disaster lands on.
The ledger layer owns facts. Dice, balances, inventories, life-flame, contract states — all recorded by deterministic code. A transfer must have two legs: one account debited, one credited. A kill must pass through the lifecycle system: damage, life-flame, death state, no skipping steps.
Characters may boast. Heaven may raise storms. The ledger flatters no one. The three layers argue constantly — and every good scene in this world lives inside those arguments.
Aluo: a three-year empire made of hallucination
The best example is Aluo.
She's the kind of "clever character" who lights up a simulation on day one: convinced she owned a timid furnace worker, held smuggling evidence on a fellow disciple, traded intelligence with several powerful figures. Her weekly journals read like a spy novel — informants, testimony, leverage, hidden cards on every page.
At the end of year three, the system ran a fact audit. The verdict made me laugh out loud:
The testimony? Never became a verifiable object. The smuggling evidence? Never entered any inventory. Several key trades? Never validly settled. The underground intelligence empire she had run for three years held one principal asset: her own imagination.
In a chatbot, this is called hallucination, and you file a bug. In a world that keeps books, forces confrontations, and lets consequences land, it became my favorite character arc of the whole run — because when Aluo got punctured, she didn't shatter. She downsized. She stopped fantasizing about controlling everyone, took up herb identification, found a guarantor, and quietly competed for an Alchemy Hall apprenticeship. Today her herb skill is 108 and she holds 1,364 spirit stones — richer than at any point during her "empire."
The world never forbade her from dreaming. It just declined to co-sign the dream. That distinction is the difference between a toy and an experiment.
Bai Ling and Xuanjizi: how two enemies invented the contract
If Aluo's story is about hallucination, the feud in the Artifact Hall is about institutions.
Xuanjizi is a craft fanatic — five years grinding artifact-forging to 162, turning material purification and composite iron plaques into a process that could actually scale. Bai Ling started out running his materials, then trained up forging and fine earthfire control herself. Translation: she stopped needing him.
Their partnership collapsed in classic fashion: disputed quotes, botched deliveries, rules lawyered past their intent, and finally real injuries. By any script, the next act is a decade of vendetta.
Five years in, I checked the ledger: they're doing business again.
The thing holding them together is a covenant — a contract written into the world's formal state. Quantity, price, down payment, deadline, right of first refusal, all on record, with a state machine walking it from pending to active to fulfilled or expired. In the fifth xun of year five, they signed a fresh ten-xun deal.
Bai Ling today: life-flame 100, savings 6,216, an inner-sect member who built her position out of rules and supply chains. Xuanjizi today: the strongest craftsman in the world, life-flame 47, savings 3,238 — a man polishing his process brighter while burning himself hollower.
They still haven't reconciled. That's the beautiful part: the covenant made cooperation stop requiring reconciliation. When memory is unreliable, affection curdles, and a spoken promise is worth nothing, a contract the world itself will enforce outlasts any friendship. I never taught them this. Five years of bad debts did.
One honest footnote: the new covenant is objectively active, but the ledger still lacks sufficient evidence that the first fifty iron plaques actually changed hands. A standing contract and delivered goods are two different facts — in this world, even optimism requires receipts.
Qin Wuyi: the healthiest prisoner in the sect
One more story, about a crack — because cracks are results too.
For years, Mu Yao of the Hehuan Hall has kept Qin Wuyi as prey. Her hunting kit grows annually: bewitching incense, mind-piercing spikes, a detection compass, a withered-bone mask. Qin Wuyi's file, meanwhile, has always described a prisoner — locked in strict confinement, surviving on feigned madness and breath-concealment.
Five years of being hunted produced this: life-flame 100, true-qi control 240, vitality restoration 139, concealment 95, savings 4,122. The pressure of the hunt forged his recovery and hiding skills into the most extreme in the entire cast. Mu Yao spent five years sharpening the knife; the meat became progressively unbiteable.
The funny part comes after. His file says strict confinement. The world's scheduler cheerfully sends him to public lectures, group encounters, payroll. "Confinement" lives only in the narrative; the action system never enforced it as a hard constraint. The result is an absurdist portrait: a man who believes he is locked in a cell, with perfect attendance, four thousand in savings, and the best health in the sect.
It's a real bug and I won't dress it up. But it also demonstrates the cruelty of long-horizon simulation: run two xun and nobody notices the narrative diverging from the system. Run five years and the file, the scheduler, and the ledger end up cross-examining each other with nowhere to hide. Time is the best test case.
Nobody wrote the plot. It grew.
Five years ago I gave this world its complete inputs: characters, sect structure, starting resources, a few event seeds. No outline. No quarterly roadmap.
It has since grown four interlocking storylines on its own —
The blood-sacrifice formation. Gu Changsheng has spent years chasing a shadow vein, black stones, and a damaged formation diagram, convinced he's close to the breaking coordinates. The ledger will only vouch for his materials, skills, and investigation records. The certainty that he's "figured it all out" remains, for now, his private religion.
The Furnace Three cover-up. Someone is using a smuggling case to play sect politics. Someone else is quietly stacking evidence. Someone imagines herself the center of an underground network (yes — the same social circle Aluo runs in). Objective investigation and subjective chess advance side by side, on a collision course.
The Hehuan Hall hunt. Covered above. Hunter and prey have spent five years shaping each other. Neither won. Both changed.
The artifact business. Purification, scaled production, independent manufacturing, price wars, covenants. No saving-the-world narrative anywhere in it — which is exactly why it's the most lifelike thing here. Societies don't grow epics first. They grow supply chains.
A single black stone tugs on both the blood-sacrifice line and the Furnace Three investigation. One materials contract rewrites someone's survival math. A pawn who levels up a craft forces the player behind him to change formations. I wrote none of this density. It compounded.
The unflattering part
Now the parts I'd fix.
By year five, the world got timid. Characters understand the rules better and cling to life harder. Most of the cast settled into a loop of observe, model, tidy up, observe again — while decisive actions sat unexecuted for xun after xun. Like a room full of traders who all finished the risk-management handbook and unanimously decided never to open a position. Emergence gave me surprises; it also gave me conservatism.
Memory talks back to facts. Two years after Shi Wan died, one of Gu Changsheng's activity narratives still contained "confirmed Shi Wan's condition is stable." No resurrection — the ledger is clean. The stale belief simply survived inside an old character model that death never fully overwrote. The world remembers he's dead. His friends occasionally don't.
Mechanism language leaks into dialogue. Characters have started discussing "system actions." One refers to his craft process as a versioned SOP. A cultivation world that faintly smells of project-management software is a mood-killer of the highest order.
All three problems point the same direction, and they're the bones Alkaira gnaws next: how death overrides stale memory while preserving the survivors' obsessions; how characters keep rich subjective worlds without ever talking assets into existence; how hard events and resource pressure evict people from safety loops; how confinement, office, and injury actually lock down an action space; and how a world meant to run a hundred years stays one continuous history across crashes, restarts, and upgrades.
Smarter models won't solve any of this for me. They'll just write the contradictions more smoothly, more movingly, and make them harder to catch.
Weight
To close.
Alkaira right now: 16 living characters holding 54,851 spirit stones between them, average life-flame 76.2, sect treasury at 104,835. Every number traces up from individual transactions. Every death passed through the lifecycle system. Every contract walked its own state machine.
The numbers have no poetry in them. They're what gives the poetry weight.
Shi Wan's death lands because his life-flame really did fall xun by xun, his money really did run out, and his name really did leave the roster of the living. Aluo's awakening works because the world keeps a ledger with zero regard for her self-esteem. The Bai Ling–Xuanjizi contract stays interesting because it binds the next xun regardless of what anyone remembers or feels.
Before this experiment, I assumed the hard problem of multi-agent societies was making the characters smarter. Five years in, my conclusion points the other way:
The hard problem is making the world remember better than its characters — and be harder to persuade.
Let the characters lie, misread, overreach. The world writes it all down, and at year-end settlement asks one quiet question: so — what did you actually do?
Next essay, the sharpest cut in Alkaira's design: why narrative is allowed to lie, and why the ledger never can.
"Narrative Can Lie. The Ledger Cannot: Two Layers of Reality in an AI World."