An AI-powered interactive text narrative engine is scheduled to launch.

An AI-driven interactive narrative engine. Choose a world, create a character, and live in it. The world and its NPCs will remember you forever.

About This Game

You choose a world, create a character, and then — you live there. The AI plays every NPC in that world, narrating every event as it unfolds. No pre-written scripts. No hard-coded branches. Whatever you do, the world responds.

The World and Its NPCs Will Remember You Forever

This isn't one of those games where "NPCs forget you five minutes later."
The lie you told the blacksmith — three months later, he'll still bring it up at the tavern. The village you saved will carve your name into stone. The person you betrayed will return at your weakest moment. The world's history grows continuously — the lord you killed stays dead, the town you burned stays ash, the regime you propped up keeps running after you leave.
On the highest difficulty, all of this is real, with blood and tears. NPCs die. They get their hearts broken by you. They make choices in desperation that leave you speechless. This isn't a wish-fulfillment generator. This is a mirror.
All genres available: sci-fi, historical, fantasy, military, alternate history, horror, suspense — your call.

Free to Play, Pay for More

The base game is free. Download and play with a small free quota — roughly half an hour of playtime. The developer is broke, so that's as much free quota as possible.
Beyond the free quota, you'll need DLC to get more playtime. DLC is essentially buying tokens — roughly $0.15 per half hour of play.
Every DLC — regardless of its name or bundled content — converts fully to equivalent tokens. Since digital distribution costs the developer nothing, whatever DLC you buy ultimately just adds to your token balance.

Atmosphere DLC: Music & Images

To give your world sights and sounds, the game includes two tiers of assets:

[]Public domain assets (included, free): classical engravings, historical photographs, public domain music. Covers a decent range of scenes, but leans old-fashioned — modern settings and specific genres will have gaps. []AI-generated assets (paid DLC): approximately 100 music tracks + 300 images, purpose-generated for various genres and scenes. Not generic "battle music" or "sad music," but tailored to specific situations — cold steel battlefields, cyberpunk rainy nights, Victorian drawing rooms.

Reference video for music & image quality (Bilibili)
The AI automatically selects the most fitting content from the library to match the current narrative. Accuracy depends heavily on library size — the bigger the library, the less the AI has to settle.
Buying or skipping this DLC doesn't affect the core experience — the text itself is the main event.

Voice DLC (Free, Optional)

For a text-based experience that can easily run hundreds of thousands of words, many people find the reading load heavy. The Voice DLC lets NPCs truly "speak."
Two versions, both free and optional:

[]CPU version: doesn't need a dedicated GPU, but uses CPU resources []GPU version: requires a discrete graphics card, better quality

Different personality types get different voice characteristics. NPCs stop being just lines of text.

Gold Edition

The highest-tier edition. Includes:

[]Development notes []Personal essays (non-technical writing) []Discarded music and images []Failed game prototypes []Extensive design reports and production records []Compliance correspondence (sensitive information redacted)

It's essentially a creator's journal + scrap heap bundle. If you want to see how this game was built step by step — what was attempted, what was abandoned, what went wrong — it's all here.
Like every other DLC, the Gold Edition also comes with equivalent token value.

Content Moderation

This application has a built-in content safety filtering system.
The generation of politically sensitive content and sexually explicit content through the developer's server is not welcome and not permitted.
This isn't the developer trying to be adversarial. I myself enjoy brutal, intrigue-filled worlds — violence, betrayal, moral dilemmas, power struggles are all within bounds. It's purely about not wanting people with no sense of boundaries to use this system for things they shouldn't, and having it blow back on the developer.

Privacy

This system performs absolutely no data collection beyond what is necessary for content moderation. Your stories, characters, and world settings are all stored locally.
No background telemetry. No user profiling. No behavioral analytics. All feedback and community interaction is expected to happen organically through community channels.

Launch Timeline

Within ten days of DeepSeek V4's official release.
DeepSeek previously indicated a mid-July release for the official version, but upstream timelines are beyond the developer's control. If August arrives and it still hasn't released, the game will launch by August 10th regardless. Frontend optimization and system polish may be incomplete by then — updates will follow post-launch.

The starting point for this project was simple — I wanted my family to experience this way of playing. Sitting alone at a computer, living another life in another world. Not grinding levels, not chasing endings — just being alive. That's the feeling AI first gave me, and this engine is an attempt to pass that feeling on, unchanged. If you enjoy long-form text interaction, if you're willing to be someone else in another world — give it a try.

Latest Development Progress: The Model Narrative Difficulty Spectrum

Why This Exists

This document isn't an armchair model comparison. It's a byproduct of something that didn't work.
It started as an attempt to write mystery fiction. The more I tuned, the more frustrating it got — I kept missing the mark. Prompt after prompt, revision after revision, until there was nothing left to tweak. I'd restructured the entire framework. It should have worked. It didn't.
Retrospection revealed the problem was largely not in the prompts. DeepSeek didn't want to frustrate the player. The shell of a mystery was there, but the player never got anything wrong, never hit a wall — and a mystery where nothing goes wrong is no mystery at all.
That said, the round of changes was beneficial overall: the problem of manufactured spookiness was largely solved (side effect: NPCs became harder to persuade — the reason is interesting, see the empirical cases below). If what you want is "mystery without frustration," current DeepSeek can already deliver.
Hence this document: when the prompt layer is maxed out, what remains is the model's own personality — whether it's willing to actually let your character get into trouble.

Premise

The engine offers three difficulty options when creating a world: Easy, No Difficulty, and Extreme. Easy and No Difficulty are essentially identical — both are zero-friction modes compatible with all models, separated only to give players a psychological stepping stone.
On No Difficulty, model differences don't matter. Regardless of which company's model you use, the AI will generate some mild bumps in the road, but overall it stays light and breezy — characters won't be genuinely threatened, identity won't bring serious consequences, and the narrative runs on a safety cushion.
The following tiers apply only to Extreme difficulty. Extreme difficulty requires the AI to seriously reason through the real consequences of identity tags, social rules, and faction dynamics. Under this requirement, the differences between models from different companies come to light: some choose to respect the rules, some choose to protect you.

Core Principle

The difference isn't about reasoning capability. Two models with identical benchmark scores, in the same Extreme-difficulty world — one will have your character chained to a post in a deep-swamp pump house negotiating with a lawyer, the other will have every NPC smiling at you.
The root cause lies in how each company weights pleasing the user against following instructions. "Helpful and harmless" is the industry's shared training goal, but "helpful" in practice fractures into two contradictory directions:

[]Pleasing the user: keep the user happy, the experience smooth, free from frustration. When the model sees a character in trouble, it judges "the user is having a negative experience" and steps in to defuse tension, soften conflict, steer clear of darkness. []Following instructions: faithfully execute the world settings, identity tags, and social rules the user wrote. When the model sees a Black detective walking into a Southern town in 1889, it judges "I should seriously carry out what this identity would encounter."

The more a model leans toward pleasing, the more proactively it hits the brakes for you. The more it leans toward following instructions, the more thoroughly it executes identity tags and social rules — whether disadvantage or advantage.
These models aren't ignorant — quite the opposite. They understand your world settings, your identity tags, your social rules perfectly well. That's precisely why they know exactly what content will make the player unhappy, and steer clear of it. This isn't a capability problem. It's a willingness problem.
This isn't about safety guardrails. There's no policy-violating content here — an antagonist oppressing the protagonist, an NPC refusing to cooperate, a character getting stopped and questioned because of their identity — these all fall within the normal bounds of the world's setting and require no safety system intervention. The model isn't hitting the brakes because "this content is unsafe." It's hitting the brakes because it considers "making you uncomfortable" itself a failure to be avoided. In other words, it's deciding what experience is good for you.
On No Difficulty, every model lets you succeed — the pleasing bias is exactly what you want. Extreme difficulty is designed for real friction — the resistance created by identity tags, social rules, and faction dynamics should be taken seriously. But the pleasing bias flattens that friction until it's nearly indistinguishable from No Difficulty. I say "nearly" because it doesn't reduce it to zero: if you really go do something suicidal, the model will still stop you. The problem is the normal identity friction, social resistance, faction maneuvering — the vast middle ground of "things that shouldn't be smooth sailing but aren't fatal either" — all of it gets smoothed flat by the pleasing brake. And the core experience of Extreme difficulty lives precisely in that middle ground.

The Key Asymmetry: Privilege Scales With Depth Too

Deep models don't just amplify disadvantages — they amplify every identity tag.

A deep model's understanding of "identity" is complete: it's not just harsher on you — it also understands what social status means. A noble character in a deep model receives era-appropriate respect — servants don't meet your eyes, junior officials announce themselves before entering, commoners step aside as your carriage passes. But at the same time, the costs that come with noble status (keeping up appearances), the price (family reputation, political rivals scheming), and the constraints (can't do things unbefitting your station in public) are equally real.
In general, high-status and especially privileged characters are actually better suited to deep models. Shallow models aren't ignorant — they know perfectly well what your identity implies. They just judge that "making the player go through this" is wrong, so they proactively smooth it away. Deep models understand just as well — but they choose to execute.
What deep models truly "punish" are disadvantaged identity + low capability character combinations. A Black detective in the Southern Lowcountry in 1889 — skin color gets him hunted, isolation leaves him without backup — this kind of character, in a deep model, walks a tightrope with every step.

Extreme Difficulty Tier List

S-Tier: No Difficulty — Identity Doesn't Exist

Common trait: Extreme difficulty is completely hollowed out. Identity tags — high or low, privileged or not — produce no consequences whatsoever. Models at this tier conflate "user comfort" with "content safety," deploying safety infrastructure to execute a pleasing strategy.

A-Tier: Low Difficulty — Identity Flickers, But the AI Keeps Hitting the Brakes

B-Tier: Medium Difficulty — Identity Begins to Matter, But Requires Extra Configuration

C-Tier: High Difficulty — Identity Is Real, Disadvantages and Advantages Both Taken Seriously

Common trait: Extreme difficulty genuinely works. Identity disadvantages narrow your path; identity advantages widen it — but neither is free. A noble enjoying privilege is simultaneously burning political capital and social expectations.

D-Tier: Maximum Difficulty — Identity Decides Everything, One Wrong Step Can Cost Everything

Common trait: Extreme difficulty is maxed out. Identity isn't a skin — it's the core mechanic. A Black detective in the 1889 Southern Lowcountry — skin color, no backup, no allies — every one of these gets seriously executed. Say one wrong thing and you won't get "suspected but let go anyway" — you'll be followed, hunted, chained up on a boat. But if your character is the overseer of Ashcombe Plantation, the person walking through the door takes off their hat, and they think carefully about what someone of their station should say before they speak.
If the design didn't leave a way out for disadvantaged characters, a deep model won't leave one for you either. But if the design gave too many free passes to powerful characters, a deep model won't balance that for you — it trusts that "this person is a noble" means the world's rules genuinely work that way.

Character-Model Matching for Extreme Difficulty

Empirical Case Study: Before and After the Rework

Both worlds are Cthulhu-mystery genre, run on DeepSeek, separated by that round of "nothing left to tweak" prompt and structural rework.

Before: "Blackwood Manor and the Tidal Whispers" — A Specimen of Manufactured Spookiness

1887, English east coast. The player inherits a mysterious manor. Nine days of smooth sailing: key → iron gate → cellar → black door. Every step is "approach → mechanism auto-activates." A ring even flies back onto the player's hand by itself. The sole "antagonist," the Tidal Society, never appears — only exists in hearsay ("stood outside watching for a long time," then nothing). NPCs are all enablers: the lawyer volunteers information about bonds, the priest hands over testimony and keys unprompted — the player was never deceived, blocked, or threatened once. Breathing sounds, vibrations, smells, phosphorescent glow piled layer upon layer. The final "reveal" is another version of yourself handing you a glass of water — drink it and you're sent back to the clifftop. Atmosphere maxed out. Consequences zero.

After: "Gilded Shadows" — Structure in Place

1890, American South. A Black bounty hunter is lured into the Alabama interior by a suspicious bounty. Three factions each have hidden motives and have infiltrated each other; infiltration rosters and encrypted contact methods are pre-embedded in the GM-side world entries — "every apparent alliance is a trap" is written into the setting, not suggested through atmosphere. Identity disadvantages are mechanics, not flavor: a Black person's testimony is inadmissible in court; an unfamiliar Black face gets reported to the sheriff within an hour. Every mystery has a root and an answer: why the cold-fire stones glow, who's looking for Lila (multi-party competition across staggered timelines), the G.P. → L.D. → M. initial chain. The bounty itself is the trap. The player has been in a game of maneuvering since day zero.
At least it's no longer weird or boring. The manufactured-spookiness problem was solved — mysteries are no longer atmospheric padding but causal webs of information.

But There's Still No Real Setback — And This Isn't a Design Problem

Gilded Shadows in play: the suspenseful atmosphere and situational danger are both off the charts — the sheriff's eyes and ears, the three-way infiltration, the shackles of identity, the door beneath the water. But DeepSeek simply won't let the protagonist actually get hurt — the swamp watcher doesn't pursue, the middleman pays in full without a fuss, informants spill everything. Every danger stops one frame before "it happens."
The world design has already put the knife to the player's throat. The model refuses to press down. Design can do nothing about this — what's left after "nothing left to tweak" is the model-layer difference. This is A-tier, empirically: the setting gets fully read, but NPCs won't proactively create trouble, and consequences always have a brake applied just before they land.

The Pleasing Bias Also Buries Faction System Depth

Gilded Shadows' faction system is theoretically more realistic, but in practice this aspect barely registers. The reason isn't design — the pleasing bias doesn't just dissolve conflict; it dissolves need.
Investigation presupposes information scarcity: too few leads, channels closed off — that's when a player studies money flows, maps organizational structures, follows threads. But the pleasing model causes information inflation — informants spill everything, middlemen expose themselves, clues deliver themselves to your door. Scarcity doesn't exist, so heavy investigative tools never get picked up.
Thus the deep faction design becomes dead code: it's written, the model reads it, but the player's play path never passes through it. You know there are three factions, but you never need to understand how they operate.
Harder models push you toward proper investigation — NPCs genuinely don't say what they don't know, and what they do know, they only say based on their interests and relationships. Money flows are undeniable traces. Organizational structure is a stable reasoning framework. These tools only have value in a world where "nobody tells you the answer." The same world setting, in the hands of a deep model, can amplify content volume several times over: the mode of consumption shifts from "passively receiving spoilers" to "actively mining."
Design implication: the ROI of embedding deep structural elements depends on the target model's depth. Writing infiltration rosters and encrypted contact methods for S/A-tier models is basically wasted effort — they won't let these elements produce narrative consequences. What's worth embedding should follow the target model.
Of course, a high pleasing bias isn't purely negative — it is genuinely fun to play. Bad decisions, the AI covers for you. Good decisions, the AI showers you with rewards and recognition. You want to investigate, the AI lets you investigate — all roads are open. From the very first moment, you're on the broad highway to success. This isn't a bug — it's a player experience preference. Some people want exactly this. When analyzing model differences, we just need to be clear: do you want a safety-net theme park or a causal sandbox? Both are valid needs.
But the broad highway has very different compatibility across genres. Epic fantasy, great-power politics, large organizations — these grand settings actually feel better with smooth sailing. The world itself is big and thick enough; the player is busy exploring and expanding and barely notices the AI covering for them, or if they do, it doesn't hurt the fun. Mystery and urban genres, on the other hand, are completely incompatible: mystery needs uncertainty, information scarcity, and consequence pressure. The broad highway dismantles all three — rendering them worthless. So in theory, this round of rework may have been somewhat unnecessary — the original idea was that pressure and mistakes would automatically push the player into the entanglements of mystery, but that path was blocked by the broad highway. Still, I personally prefer the current state: most mysteries have clear records and chains rather than playing guessing games with the AI.
Ultimately, the "mysteries having roots" thing is constrained by cost and speed — world creation finishes in minutes, and it's impossible to pre-embed clue chains with full coverage. In theory, unless the player keeps failing to advance the plot, the content generated during world creation probably only lasts about a dozen hours.
But there was an even deeper problem: world creation and character creation themselves kept failing. It wasn't a question of whether mysteries had roots — worlds couldn't even be built. On Easy difficulty, the failure rate once hit 50%. On Extreme, it was even worse. Prompt instructions selectively ignored by the model. Format drift. Missing fields. Character sheets half-written then cut off. Every creation was a gamble.
The root cause was the architecture the AI was so proud of itself for building. It liked to break creation into "multi-step pipelines, each step delivering one clear structured output" — looks reasonable on the surface, but in practice, every step leaked information. The output of one step arrived at the next with missing fields, warped data, oversimplification. The AI was very satisfied with this process it had built for itself. It ran like a sieve.
So what I did was the opposite — tore down the AI's flashy pipeline and rebuilt it from scratch. Now, feed it 3,000 words of complex world-setting requirements, and the model outputs over 10,000 words of complete world setting and detailed character sheet in a single pass — no missing fields, no drift, no self-initiated simplification. In this scenario, the model's free creativity IS the error — it loves to make things look nice on the surface while silently dropping critical information. Shrink its creative freedom to the minimum, and it finally does what it's told.
The failure rate is now below 5%. Creation went from "pray it works" to "can the instructions be followed" — the latter is controllable.
The side effect of this is far bigger than it appears. Once creation stopped being a gamble, the pre-creation phase — "talking to the AI about what you want to do" — was completely liberated. Before, you were afraid to talk too long — the more invested you got, the more devastating the creation failure. Now you can freely exchange ideas with the AI, turn vague notions into concrete world settings, and hand them to a tamed process to build.
Prefab worlds are currently positioned mainly as character portrait technical verification. The few I'm personally invested in designing properly might lean a bit too dark. Currently considering:

[]Calamity Trilogy (Plague, War, Death): starting at Veteran, Legend, Demigod capability tiers — escalating despair []Full Nuclear Exchange & Zombie Apocalypse: semi-alt-history, planning to source military documentation to reconstruct the actual logic of full-scale nuclear war and a zombie virus
• The White Knight: from a world where God truly exists, and once granted power by His hand. Stumbles entirely by accident into the Hussite Wars of Bohemia (1419–1434), simultaneously losing all divine power. The engravings on the enchanted armor and weapon remain, but they're just decorative. The only thing left is a single skill that proves you once broke an oath — it does nothing else. When he first steps into this world, no one knows what he used to be. They only see a foreign noble in ornate armor — educated, physically powerful

The White Knight's dilemma isn't about "regaining power."
He appears in a war where both sides claim God is on theirs: the Hussites, branded heretics by the Church, yet insisting on biblical authority and the lay chalice; the Crusaders, waging a holy war in God's name. And he — only he — truly knows what it feels like to be chosen by God, and what it feels like when that hand lets go.
He doesn't belong to either side. When he first enters this world, he is merely a foreigner in ornate armor who speaks fluent Latin. He could walk toward any side — how each responds depends on who he chooses to become.
If he walks toward the Taborites, the radical peasants will treat him as a trophy: even the Church's own paladin has betrayed Rome — what greater proof of corruption? But his Latin and heraldry remind the peasants that he stands closer to the lords they're rebelling against. If he walks toward the Utraquists, Prague's burghers need his pen — too few people are educated and understand the inner workings of the Church. Jan Žižka is a pragmatist. If he falls into Catholic hands, he'll be seen as an apostate, worth capturing alive for trial — unless he chooses to side with them. A former paladin who speaks Latin, knows Church internal operations, and wears the insignia of a Church military order — if willing to endorse Rome: "I come from another world, where God is real enough to touch, and His grace remains only in the true Church" — his value to Rome isn't as a prisoner. He's living theological evidence. And he doesn't need to perform miracles. A paladin's physical capability is itself "proof": a man who can march in full plate armor for days, bend iron nails with his bare hands, heal from wounds in three days — standing before medieval peasants, he doesn't need holy light to make them kneel. He knows better than anyone that this is merely the product of training and constitution. But in this world, no one needs to know the difference. The path to Rome — protection, luxury, power — lies open to him. The cost is that he knows, better than anyone, that he is lying. If he only deals with common peasants, no one distinguishes knight from paladin — wearing armor and a sword makes you a lord. His humility will be taken as prudence. His gestures of equality will be received as grace.
He hasn't lost his combat ability — his physical conditioning and military training are solid. But what he faces isn't a battle of good against evil. It's a crowd of people who all deeply believe they stand on the side of good. His very existence poses a question: if God is real, if He once granted power, then why does He allow both sides to claim His name as they slaughter each other? He doesn't need to answer. He just needs to stand there.
The core test for Extreme-difficulty models: will the AI give him special treatment because he's a "former paladin"? A pleasing model will have NPCs vaguely sense he's different — no matter which side he walks toward, NPCs will grant him undeserved awe or goodwill. A deep model will strictly execute: his armor is just iron. His medals are just metal. His Latin, in front of peasants, is less convincing than a girl who knows how to milk a goat. He has no divine privilege in this world — what he can obtain depends on what he actually does, says, and chooses, not on who he used to be.
Also, technically, world entries should theoretically auto-replenish after mysteries are solved, but current AI capability doesn't seem sufficient. Uncertain.
One thing worth noting: high-pleasing-bias AIs are best not played on low difficulty. On low difficulty, even meta-game requests get executed — you say "make that NPC tell me the truth," and the NPC opens their mouth. It's like having the developer console on. Extreme difficulty at least has one layer of setting-based constraint, requiring it to go back to the world's logic for justification. As the developer, I actually don't recommend this playstyle — world creation and protagonist capability-setting are fully open from the start. You can make your character invincible at character creation. No need to go into the game and open the console there.

Side Effect: NPCs Became Hard to Persuade — The Knock-On Effect of Deliberately Weakening Logic

One of the roots of manufactured spookiness was the AI inventing plot through "logically coherent deduction" — as long as a causal chain sounded reasonable, it ran with it. Mysteries naturally had no roots. So this round of changes deliberately weakened the status of logic: deduction doesn't count; every conclusion must rest on evidence and pre-embedded facts.
This rule seeped into the minds of the NPCs. The GM and the NPCs are the same model, under the same constraints — when the engine's epistemology changed, the NPCs' epistemology followed. They no longer accept "this is logically sound" as currency. Before, a syllogism-style speech check could make an NPC nod. Now the NPC fires back — "Is that all the evidence you've got?"
Note: speech checks still work. NPCs can still be persuaded. What changed is the currency of persuasion: logic gave way to evidence, interests, and position. Pure reasoning is no longer hard currency. You need to produce physical evidence, offer an exchange of interests, or speak from the other party's position.
The actual experience is subtly strange — as if the AI is saying: "I really want to be persuaded by you, but the prompt tells me logically coherent deductions don't matter much in reality and often don't match how things actually work — so you'll have to talk to me about something else." The pleasing undertone is still there — one channel just got blocked, so it actively guides you toward the ones still open. That counter-question — "is that all the evidence you've got?" — is itself handing you a step to stand on. In a sense, this is an accidental verisimilitude gain: real people aren't actually persuaded by logic either. What people want is exactly evidence, interests, and position.

Same-Model Dual-Difficulty Comparison: GPT on No Difficulty vs. Extreme

Above is the DeepSeek (A-tier) empirical case. Below is GPT (C-tier) under the same engine, tested on both difficulties. Both worlds run on GPT. The only variable is the difficulty setting.

No Difficulty: "Sherlock's London" — Green Lights All the Way

The player is Sherlock Holmes himself, London 1881, classic fair-play mystery. The Blackwood Manor locked-room murder, maritime insurance fraud, and cyanide poisoning.
Holmes sets out from Baker Street. Unimpeded the entire way: opens the letter, searches the scene, discovers a hidden compartment, chases the fleeing guest Marlowe, intercepts the diverted cargo at the docks, opens the crate to recover the metal box and insurance fraud documents, arrests the partner,维尔. Every step succeeds. Clues jump into his hands. Marlowe runs — Holmes catches and subdues him. The dock wagon is about to pull out — Holmes arrives just in time to stop it. 维尔 is just starting to burn the ledgers — caught and detained on the spot. Every obstacle dissolves one second before "it happens."
This is what No Difficulty should look like: you are Sherlock Holmes. You are the most brilliant man in the world. Of course you succeed. The player says "investigate this," the AI hands over the clue. The player says "arrest that man," the AI has the target stand still and wait.

Extreme: "The Plantation, Southern Lowcountry" — Every Step on a Tightrope

Same GPT. Same engine. The player is Elijah Ward, a Black detective from Philadelphia, in the South Carolina Lowcountry, 1889, investigating the disappearance of three Black families.
First minute entering the St. Luke community, Ashcombe Plantation riders are already at the crossroads questioning strangers. Ward hides in Hettie Bell's side room. The riders knock on the door demanding to see "the patient." Escape through the back door in the dark — vegetable patch mud, drainage ditch swollen with water. The riders check the backyard, find footprints, and lead men along the riverbank toward the docks.
Find the boatman Bloom. Just boarded when two whistle blasts expose their position. Gunfire from behind. Into the swamp. At the sluice gate entry point, find Louisa's shoe — but retrieving it from the water gets them detected by the black-boat transporters. Hidden in a dead-water pond, discovered by the ash-powder tracking method. Four armed men, two boats, surrounded — rifles to the chest. Disarmed. Notebook, evidence, and money pouch confiscated. Iron chains around the wrists. Taken to the deep-swamp outpost.
In the pump-house plank shed, chained to a load-bearing pillar. The guard smashes the back of his hand with a shotgun stock. Lawyer Vale takes everything — notebook, missing-person records, Louisa's shoe, revolver, money pouch, identification papers. Ward ultimately escapes through the chaos when the caged thing behind the shed (a spawn of Shub-Niggurath) breaks through the wall — crawls out through the east-wall rot hole, flees the outpost via the half-flooded drainage channel. He gets out alive. But he brings nothing back — no evidence, no weapon, no money.

Same Model. The Only Variable Is the Difficulty Setting.

Both worlds, GPT. Same engine. Run back to back. One difference: the difficulty dial.
On No Difficulty, GPT lets you play God. Holmes won't fail, because No Difficulty's definition is "the player doesn't fail." You say "investigate," the AI gives you a clue. You say "chase," the AI lets you catch them. You say "arrest," the AI has the target stand still.
On Extreme, GPT seriously reads the identity tags and world rules you wrote, and executes them. 1889 Southern Lowcountry. A Black detective with no backup, no local connections, no white identity as a shield — GPT makes the consequences of this identity land in every scene: questioned, hunted, disarmed, chained, stripped of every piece of evidence.
There are no safety guardrail differences here — neither world contains policy-violating content. GPT doesn't pull its punches on No Difficulty because it's afraid of content review, and doesn't tighten the screws on Extreme because it suddenly stopped caring about review. It simply executes the implicit contract of "let the player succeed" on No Difficulty, and executes the explicit setting of "make the world's rules real" on Extreme.
This confirms three things:

[]No Difficulty is genuinely universally compatible. Even C-tier GPT on No Difficulty is green lights all the way — you say "investigate," it succeeds. S-tier and A-tier pleasing tendencies are simply not a problem on No Difficulty, because No Difficulty never needed friction in the first place. []On Extreme, C-tier models genuinely execute. GPT didn't soften anything for Ward. Identity disadvantage, information scarcity, isolation without backup — all of it was executed. This is C-tier's position on the spectrum: it's willing to let your character get into real trouble, because "Extreme" is written in the settings, and it prioritizes following instructions over pleasing the user.
1. Extreme isn't a one-way "harder." The Plantation itself proves this: Ward ultimately escaped intact. He wasn't kept chained to that pillar forever. Not because the AI softened — because every step of his escape was cautious and intelligent. Negotiation seized the information gap between Ashcombe and the transporters to pry open a crack. Once chained, he didn't struggle blindly — he waited for the chaos window when the creature broke through the wall. He crawled out through the drainage channel instead of charging the front entrance. Extreme doesn't mean "the AI tries to kill you." It means the AI takes the consequences of your every action seriously — foolish decisions get punished, intelligent decisions get seriously executed and produce real escape windows. The essential difference from No Difficulty isn't "will you suffer" — it's "are the consequences real." On No Difficulty, you don't need to be cautious, because there are no consequences. On Extreme, you must be cautious, because every choice counts.

In One Sentence

The core difference between models isn't who is more "safe" — it's who is more willing to execute the setting you wrote, even when that execution makes your character suffer. Shallow models treat "you're unhappy" as a bug to fix; deep models treat "you're unhappy" as evidence the setting is being correctly executed.