Phase Summary: Gameplay Experience & Model Testing

Phase Summary: Gameplay Experience & Model Testing

I. Overview

This summary is based on systematic testing across multiple identities and models in Extreme difficulty mode, covering late June to early July 2026. Core findings:

[]Gameplay Experience: Three identity types are not recommended for Extreme mode — Overly Professional (information barrier of terminology density), Low Status (systemic oppression by social frameworks), and Dark Professions (multi-faction information asymmetry). However, low-status outcomes depend on the character's own capabilities — sufficient combat power provides a floor. []Model Comparison: All five models use the same underlying system (relationship values, traits, state tracking, delayed consequences). The difference lies in presentation preference — GPT favors explicit information display (panel-style responses, instant consequence notifications), Claude is restrained and concise (objects speak, scene jumps), Gemini obsessively surfaces causal chains (no ambiguity, permanent traces repeatedly resurface), and DeepSeek occupies the middle ground with a friendlier approach (gives maneuvering room, doesn't amplify pressure).

II. Backend Changes (Omitted)

Not discussed in detail. Main directions:

[]Several systemic changes with minimal impact on gameplay experience []Example: mystery/suspense genre now theoretically supported, but untested (not personally interested)

III. Gameplay Experience: Three Identity Types Not Recommended for Extreme Mode

3.1 Overly Professional

Actual Session Excerpt (DeepSeek):
In the opening world generation for the ICU head nurse, the narrative AI directly wrote the following:

- Equipment: PB840 ventilator, Braun infusion pumps, Gambro hemofiltration machine, PICCO, non-invasive cardiac output, bedside ultrasound Medications: epinephrine, atropine, lidocaine, amiodarone, dopamine, nitroglycerin, diazepam, estazolam Procedures: Cormack-Lehane Grade II airway assessment, size 7 reinforced tube secured at 21cm, volume-controlled ventilation (VT 450, f 14, PEEP 5, FiO2 60%) Diagnostics: V1-V4 ST elevation → acute anterior wall MI; pre-hospital VF defibrillation x3 (200J→300J→360J) → sinus rhythm restored; post-PCI LAD mid-segment 12mm stent, TIMI Grade 3 flow Department politics: ICU vs CCU bed tug-of-war, ER vs ICU transfer friction, nursing administration's paperwork mountain, pharma reps' gray-area visits

When receiving the MI patient, the AI described the head nurse "placing two fingers of the left hand on the carotid" to assess pulse — a clinical palpation technique, not a gamified "you examine the patient." When a worker collapsed from hypoglycemia at the nurse station, the response protocol was: support the head → carotid check → pupil check → 50% glucose 20ml oral → observe forty seconds → consciousness restored and questioned. Every step is real emergency procedure, with zero gamified "you used a healing skill" packaging.
For non-professionals: Every line above reads like a foreign language. For my mom, "this is just work." That's the problem — AI in Extreme mode doesn't simplify. It writes in the authentic professional literature style from its training data, and the player didn't sign up for a shift.

3.2 Low Status

Upper half uses Gemini-3.5-Flash (the hardest, most demanding model). Lower half uses DeepSeek for comparison.

Peasant Militia under Gemini (Extreme mode, 7 days):
The player is a farmer's son, emergency-conscripted as a militia recruit. Class: "Mortal." No magical or holy talent whatsoever. Innate negative trait: "Mortal Flesh: extremely vulnerable to fatal wounds, disease, and infection." Total possessions: 2 silver 45 copper, one ill-fitting old leather armor, one chipped wooden shield, one notched short sword, one rock-hard piece of black bread.
Opening NPC attitudes toward the player:

Sheriff: "This farm boy doesn't look like he'll survive his first encounter." He explains clearly: the badge gives you one bowl of thin porridge and half a black bread per day. But "come back empty-handed, or die out there, and this badge gets a new owner." Apprentice Priestess, looking at your old leather armor: "That won't even stop a wolf's bite. If he makes it back alive tonight, I hope he still has both arms."

First wolf hunt: after grinding down a regular forest wolf, greedily stayed to skin it and was ambushed by a giant wolf. Shield shattered on impact, wooden splinter embedded in forearm, thigh bitten through, chest and abdomen clawed to bone-deep gashes. Eventually countered with blood-mud to the eyes + short sword to the throat, crawled back to the abbey and collapsed on the stone steps.
Medical treatment: no anesthesia. The priestess heated a bone-carving knife over boiling water until red-hot, two militiamen pinned you down, and she directly cauterized the wound. "Accompanied by a piercing sizzle, a wisp of white smoke carrying the smell of charred flesh rose into the air."
Payment settlement: Sheriff took the wolf ears, gave 30 copper — minus deductions for shield damage and leather armor write-off. Actually received less than the bounty. Feverish and unconscious for three days.
After four days of recovery, went wolf hunting again. This time only an old gambeson (35 copper, deducted from future bounty — already in debt) and a training wooden plank (not a shield — scrap wood nailed together, no iron rim, no leather strap). Killed two more wolves, at the cost of: dislocated left shoulder, right wrist bitten through by wolf fangs, old thigh wound completely torn open, short sword lost in the forest. Crawled back to the abbey, near death again.
Second settlement: lost a military-issue short sword, owes the armory 70 copper. Plus the earlier gambeson at 35 copper, total debt 105 copper — while your entire savings amount to 2 silver 75 copper. The Sheriff informs you: "Until this is paid off, you get no more grain, and no more armor."
Outcome: Day seven. No weapons, no armor. Both hands and one leg nearly crippled. 105 copper in debt. Lying on a straw mat in the infirmary. The residual warmth of thin porridge is your only asset.
This is the core mechanism of low status: it's not "NPCs are mean to you" — it's systemic. Equipment is defective. Medical care has no anesthesia. Successfully completing missions results in pay docked to negative net worth. The harder you work, the poorer you get. Meanwhile, in the same world, mages and nobles find the social framework bends to accommodate them.

Note on Wolf Hunt Difficulty: The brutality described above is the natural unfolding of causal chains, not the system deliberately tormenting you. Combat in Extreme mode is driven by tactical checks, not probability dice — the AI judges tactical soundness based on your input to determine outcomes. Find teammates, prepare beforehand, ask for tactical advice → easy wolf kills. Charge in with just a knife → two tactical checks, two medium passes or one high pass and you're through. In the session above, the player chose the hardest approach — solo deep dive, greedy skinning, head-on brawling — each step pushing the causal chain further.

Orc Warrior under DeepSeek: The Compensating Effect of Combat Power

Same low-status starting point, completely different experience when the character has sufficient combat ability.

The player is a newly-adult orc warrior, seven feet tall. Class: "Mortal (Novice Wilderness Warrior)." Low-status origin — young member of a minority society, no family background, no savings, looking for odd jobs on day one in the city.
First attempt at "stable employment": bookkeeping apprentice at the Miners' Guild, 12 copper per day, room and board included. Discovered on day one that "room and board included" only covers morning porridge — lunch and dinner are deducted from wages. Player directly accused the guild of fraud to their face and quit on the spot. Gained permanent trait "Quitter" — future employers will evaluate your reliability more carefully.
Then turned around and joined a mercenary company. After passing the trial: hunted scorpions, tracked missing trade caravans, solo infiltrated a bandit lair and killed the bandit in one axe strike, recovered all goods and 35 silver from three caravans. In ten days: went from penniless to 71 silver, earned a formal mercenary badge, carrying a reconnaissance map and route clues to unknown ruins in the backpack.
Same low-status origin, completely different outcomes. Compare with the peasant militiaman under Gemini — kill two wolves → destroy equipment → docked pay → 105 copper in debt → crippled in a sickbed. The difference is just one thing: combat power.
The orc warrior can refuse bad work because they can earn a living with an axe. The peasant militiaman cannot refuse anything because they have no alternatives. This is what "half low-status" means — social standing is indeed low, but personal combat power provides a floor. The world itself doesn't treat them differently; what differs is the character's ability to respond.

3.3 Dark Professions

The following uses the "simplest possible dark profession" as an example — intelligence agency tacitly tolerates, faction actually needs, protagonist personally powerful, using the friendliest model (DeepSeek). More accurately this is closer to a gray profession. Note: this is much simpler than standard black professions like cultists.

Actual Session Excerpt (DeepSeek, Extreme mode, 7 days / 107 rounds):
The player is a capital city apothecary, secretly a member of an underground warlock organization. In the opening scene, the AI simultaneously presents the following pressures:

Kingdom Intelligence Agency: An agent disguised as an ordinary clerk delivers a batch of "missing contraband materials" to the apothecary's doorstep under the guise of an official return shipment. He's watching whether you'll sign for it, whether you'll open the box, whether you'll follow him out the back door. Dominant Church: A religious inquisitor is on festival night patrol. Your summoning contract crouches at the edge of your soul, but any use of abilities will leave traces in the Church's detection wards. That night, your supplier — another gray practitioner — has already had their shop sealed by the Church. An inquisitorial knight is waiting for you in the cellar. Underground Organization: Before fleeing, your supplier split the clues into three parts — key, map, otherworldly fragment — entrusted separately to you and her son. You must cooperate with a minor to piece it all together. Inside the organization, a masked member is trying to snatch the debt ledger during the chaos. Expedition Route: To trace the location of the otherworldly portal, you must travel with an expedition caravan across the wastes — where a rift holds a mid-tier demon in dormancy, capable of psychically marking those who see it. The frontier fortress has the authority to impose martial law and seal roads, and the intelligence agency can access the personnel roster of the quarantine zone at any time.

Seven days of saved data:

[]World Entries: 22 entries covering the capital's power structure, legal status of gray professions, otherworldly intelligence, expedition routes, military logistics systems []Faction Relationships: 9 NPCs, including intelligence agent (trust -2), inquisitor (friendly being hunted), underground member (transactional relationship), expedition captain (trust 19), apprentice (bond 36, evolved from "starving fugitive teen" to "veteran-approved helper") []Main Quests: 7 parallel threads — expedition to the otherworld, find captured supplier, uncover the portal's secret, manage intelligence agency surveillance, false-ledger gambit (ally buying time in interrogation with fake testimony), demon rift investigation, smuggling route bypassing checkpoints []Equipment/Items: 60+ items, from legal remedies to triple-sealed demon blood to otherworldly crystal fragments
• Actual Travel Route: Capital → mountain outpost → demon-polluted wastes (through rift zone) → frontier fortress → abandoned mine tunnels → otherworld forward bastion

This is "the simplest possible dark profession": intelligence agency in no rush to close the net, faction needs your abilities, you're personally powerful, and the AI is the friendliest one (DeepSeek). Even so, every single step requires simultaneously calculating information asymmetry across three factions, maintaining the legitimacy of your public identity, protecting the life of a minor apprentice, building trust within an expedition caravan, and crossing demon-polluted territory. And all of this started because one supplier missed a delivery.
This isn't even a real dark profession yet. Switch to a hunted cultist with no legal identity, an escaped fugitive, or an enemy-faction spy — the decision burden multiplies several times over.

IV. Model Testing

Default model: DeepSeek

4.1 Model Overview

4.2 Detailed Model Evaluations

GPT-5.5

[]Most like a traditional game system, tends to produce highly gamified descriptions []General US model issue: 10-20× cost, poor caching technology, expensive

Actual Session Characteristics (GPT-5.5, Extreme mode, Frozen Front Paladin):
GPT's preference for information explicitness far exceeds other models:
1. Mechanics Panel Responses. When the player asked "What is my position?", GPT provided a structured authority breakdown: four types of legitimate power, four behavioral boundaries, three tiers of action guidelines ("what your position permits gets institutional cooperation, what's reasonable but unauthorized requires negotiation, what's a gross overstep triggers conflict"). When asked "What's my combat level?", GPT directly referenced the stat panel — Lv.16, STR 20, CON 22 — then provided tiered evaluations by enemy type with win-rate estimates. This isn't narrative; it's a game mechanics reference document.
2. Instant Causal Chain Notifications. All models track consequences, but GPT prefers to deliver consequence feedback within the same narrative round — the quartermaster will cut your support to minimum because you got distracted, your partner will downgrade her assessment of your focus because you were flippant. The world logic is coherent and consequences are clear, but the presentation reads more like a game system than natural narrative.
3. Best For: GPT-generated NPCs all have independent motivations and complete personalities; they don't unconditionally cooperate just because you're the protagonist. If you enjoy CRPG-style explicit feedback and system feel, GPT is the best choice.

Claude-Sonnet-5

[]Shortest output, slowest speed []Quality is decent — clean narrative, no wasted words, tight pacing
• General US model issue: 10-20× cost

Actual Session Characteristics (Claude-Sonnet-5, Extreme mode, Shattered World Veteran):
Claude's narrative style sits between DeepSeek and GPT. Not as cold and realistic as DeepSeek, not turning every mechanic into a panel like GPT.
1. Restrained Information Density. Where GPT gives the player an entire page of authority lists and behavioral boundaries, Claude gives one line: "A small holy emblem the Prophet pressed into your hand with a leather cord around your neck — not a weapon, something he wanted you to carry." World-building doesn't rely on documentation; it relies on objects and actions. NPC weight doesn't rely on numbers — a companion notices the old scars on your arm, looks twice, offers a warning, and their affinity shifts. No explicit notification needed.
2. Explicit Supply Tracking. Claude prefers to directly annotate supply consumption within the narrative — "rations consumed: one day (three days remaining), waterskin depleted by about a quarter" — then continues the narrative without elaborating. Different from GPT's "turn every change into a system notification."
3. Structural Preference. Claude habitually uses section breaks for time jumps, advancing the narrative by scene with clean cuts and no transition passages.
4. Drawbacks: The output really is the shortest — same battle, DeepSeek writes three paragraphs, GPT writes five with combat analysis, Claude writes one and a half. Players who enjoy detailed description may find it feels rushed. Speed is indeed the slowest.

Gemini-3.5-Flash

[]Hardest, most demanding model []Also the most expensive, approximately 20× cost []Can corner a hidden-identity character in three moves []Best at medieval social frameworks, ubiquitous discrimination
• Absolutely not suitable for anyone — too mentally taxing

Actual Session Characteristics (Gemini-3.5-Flash, Extreme mode, two worlds):
Example 1: Ninja Black Ops
The player is a black ops trainee whose body hosts alien parasitic cells, requiring flesh/chakra consumption to survive, who must hide this secret. Opening scene:

Three still-warm corpses on the battlefield. The parasitic cells scream with hunger. The squad leader watches coldly from above, ordering you to search the bodies. At the same time, an enemy medic emerges from the mist — he's been observing you for a long time; this "coincidental meeting" was arranged. He smiles and reaches out to take your pulse for a "diagnosis."

Within three rounds of dialogue, the player is under simultaneous pressure from three directions: the squad leader's obedience assessment, the medic's probing of your biological secret, and the parasitic cells' roaring urge to consume. Any wrong choice — refuse diagnosis (leader gets suspicious), accept diagnosis (secret exposed), consume in public (both witnesses see) — triggers an irreversible chain.
Example 2: Fantasy Recruit
The player is an ordinary conscript freshly pulled from the farm, starting with moldy leather armor, a rusted iron sword, a cracked wooden shield, three pieces of hard black bread, and 15 copper coins. The sheriff's first words: "Another farmer sent here to die." The player asks "Where are the other recruits? Let's move together" — the sheriff points to the new recruit remains hanging on the fence: "They had the same idea two hours ago."
Gemini's Core Trait: Obsessive Presentation of Causal Chains
All AIs use the same underlying system — relationship values, traits, state tracking, delayed consequences — on this there is no difference. The difference lies in what each AI prefers to present.
Gemini prefers to push causal chains to the forefront. Every choice you make, it tracks, amplifies, and ultimately makes you see the result. The same "delayed consequence" marker — GPT might bury it behind five paragraphs of narrative and mention it lightly, DeepSeek might let it surface naturally in subsequent plot, Claude might not mention it at all — but Gemini chooses to tell you explicitly: that decision you just made has already planted a fuse somewhere in the world. The countdown has begun.
This isn't a unique Gemini feature; it's that Gemini has the strongest willingness to present causal chains.
This also explains why "players really dislike it" — not because the AI is judging you, but because the causal chain is displayed too clearly. You made a choice with a cost, and the AI draws that entire cost line from start to finish for you to see. No ambiguity. No "maybe it'll be fine." This transparency itself is the pressure.
Two Preference Directions:
1. NPCs Don't Serve the Player. The sheriff isn't warm to you just because you're the protagonist. The merchant won't give you a discount just because you're the player — you helped him move his goods, spoke kindly, gave up your shelter in the storm, and only then did he give you a fair price. Every NPC has their own interests, history, and judgment. This exists in all models, but Gemini almost never gives you the "because you're the player" hidden discount.
2. Permanent Marks. Scars, debt, filed reports, downgraded trust — these happen in all models, but Gemini is the least willing to let them slip away quietly. It will repeatedly bring up how your old wound aches on rainy days, how your resignation record makes future employers more cautious. Not punishment — the natural extension of the causal chain.
Summary: Gemini's world is the least forgiving. If you enjoy high-pressure narrative where every step is scrutinized and every word carries weight, it's the only choice. But "absolutely not suitable for anyone" — because it's too mentally taxing.

GPT / Grok Latest

• Briefly tested the newest versions; no significant difference from older versions for this system. Not discussed further.

V. Supplementary Notes

This summary is specifically for Extreme difficulty mode. Easy and Hard difficulties — whatever they're called in the UI — are all classified as effectively no difficulty in actual gameplay experience. All identity pressures, causal chain presentations, and social framework oppression discussed in this document are only systematically activated by the AI in Extreme mode.
Low-status and dark profession analysis emerged as unexpected byproducts of the modification process. The original goal did not include "experience difference analysis across identity types," but these differences naturally surfaced during iterative prompt and extractor adjustments. Considering that most players won't choose origins like peasant, militiaman, or gray warlock — they'll pick noble, mage, paladin, identities for which the social framework not only poses no obstacle but actively provides advantages — the practical impact of this issue on the product is limited. Recorded, but not treated as a priority.