bokuraMachi

How a World Begins

Every playthrough starts with two people: Adam and Eve. They are templates, not fully random, but slightly randomized each run in personality, tendencies, and starting knowledge. That small initial randomness is intentional. It creates butterfly effect divergence — two worlds that start almost identically will diverge completely over generations because tiny differences compound. No two playthroughs produce the same civilization.

Language Architecture

Tier 2 — Weighted Probabilistic Grammar: Not template sentences. Not a flat phrase list. A structured grammar system with category slots, e.g.: [subject_ref] + [feeling_verb] + [emotion_word] → “i've been feeling lonely.”

Each NPC has weight distributions over each word category: Feeling verbs (~50 entries), Emotion words (~100 entries), Action words (~80 entries), Subject references, time markers, qualifiers, etc. A sad NPC has high weights on negative emotion words. A farmer has high weights on harvest/labor action words. Individuality lives in the weights, not the words.

Word categories are separated by grammatical role so combinations stay coherent. A lightweight coherence layer (small hidden state) tracks topic context to prevent incoherent fragment chaining.

Inheritance: Children inherit a subset of parent weight distributions, biased toward fragments the parent uses most, then drift slightly from their own experiences.

Mutation: Small random drift occurs during inheritance — a word swaps, a phrase shifts. This preserves butterfly effect divergence across generations and prevents cultural convergence toward the player's dialect.

Tier 1 — Small LMs: Real generative language models. Complexity scales with active interaction hours (focused, direct interaction time, not just proximity):

• 0 – 0.5h — No LM yet, tier 2 only, feels like a stranger
• 0.5 – 2h — Bootstrapped LM, ~500 tokens, surface-level acquaintance
• 2 – 8h — Small LM, ~1500 tokens, consistent personality
• 8 – 25h — Medium LM, ~3000 tokens, remembers things, feels real
• 25h+ — Full tier 1, ~5000 tokens, genuine unpredictability

Vocab ceiling at 5000 tokens is intentional. Returns diminish beyond that. LMs grow incrementally each session, not in sudden jumps.

Bootstrap Handoff: When an NPC is promoted from Tier 2 to Tier 1, their LM is bootstrapped from their existing Tier 2 weight distribution. The handoff must feel seamless — the character should not noticeably “change” the moment they become a close friend.

Inheritance for Tier 1 Children: A child of a Tier 1 NPC starts with a richer weight distribution and a head start on LM development. A child of a distant or culturally foreign NPC starts with weights that have little overlap with the player's language — they will feel hard to talk to. This is intentional and emergent, not scripted.

The NPC Brain — Full Architecture

Every citizen has an AI brain. The language system is one component. The full brain consists of:

• Emotion Model — Tracks and generates emotional states. Affects behavior, speech, and responses. Emotions influence decisions, not cosmetic.
• Language Model — Tier 2: weighted grammar system. Tier 1: small trainable LM. NPCs learn language from people around them. Children learn from parents. Speech patterns and vocabulary propagate socially.
• Senses Model — Processes what the NPC perceives: who is nearby, what is happening, environmental context. Feeds into decisions and emotional responses.
• Memory System — NPCs remember past interactions, relationships, events. An NPC who was hurt by someone responds differently to them than to a stranger.

All four systems interact. An NPC sees something (senses), feels something (emotion), recalls relevant history (memory), and responds in their own voice (language).

Generational Learning — How Culture Emerges

Parents teach children. Children learn from parents and from everyone they spend time with. Knowledge, behavior, speech patterns, and values propagate through the population organically. No culture is scripted. Every civilization emerges from the specific chain of interactions that happened. A child raised by two cautious parents in an isolated part of town will be a different person than a child raised by adventurous parents in the social center.

Over generations, genuine cultural drift happens. The civilization the player encounters at generation 5 is a product of everything that came before it. The mutation system ensures this drift never fully converges, even across many generations.

The Compute System

Relationship depth maps to computational depth: Tier 1 (active, close) — LM hot in memory, 1–2 max simultaneously. Tier 2 (active, acquaintance) — Grammar sampler running, cheap. Tier 3 (frozen) — Paused on disk, no CPU cost.

Fast-Forward Catchup: When approaching an NPC not seen in a while, a loading screen runs their simulation forward at high speed, catching them up on what they would have experienced since last contact. The loading screen is doing real work. For Tier 1 NPCs at high interaction hours, this catchup is heavier. Expected time on mid-range laptop hardware: 2–10 seconds for most cases.

Promotion / Demotion: NPCs move between tiers based on active interaction hours. Tier 1 cap is 8. Hitting the cap forces a demotion choice — who do you let drift? This is emergent social pressure, not artificial game design.

Target Hardware & Town Size

Target platform: mid-range laptop, no dedicated GPU required at target scale. Reference hardware: quad-core CPU (e.g. Intel i5/i7 8th gen or later), 16GB RAM, integrated graphics. Recommended town size: 150 citizens.

• 20 citizens — 3 T1, 10 T2, 7 T3 — 4GB RAM, any modern CPU
• 50 citizens — 5 T1, 20 T2, 25 T3 — 6GB RAM, quad core
• 100 citizens — 8 T1, 30 T2, 62 T3 — 8GB RAM, i5 8th gen or later
• 250 citizens — 8 T1, 60 T2, 182 T3 — 12GB RAM, i5 10th gen or later
• 500 citizens — 8 T1, 80 T2, 412 T3 — 16GB RAM, i7 + GPU helps
• 1000 citizens — 8 T1, 100 T2, 892 T3 — 24GB RAM, i7 + 4GB VRAM GPU recommended

Note: Tier 1 cap stays at 8 regardless of town size. 1000-citizen towns lose the intimacy that makes bokuraMachi interesting.

What the Player Does

You live in the town. You talk to people, build relationships, exist. The world changes around you whether you engage or not. You influence the civilization by being part of it — the relationships you form, the things you teach people, the interactions you have all feed into the simulation. There is no win condition. There is no lose condition. The game ends when you stop finding it interesting, or never.

Key Design Principles

• Everything emergent, nothing scripted. If you're considering a scripted event or hardcoded behavior, ask first whether the simulation can produce it naturally.
• Parsimonious complexity. Simple rules, deep behavior. Brain components are individually simple. Emergence comes from their interaction over time.
• The player is not special. Any mechanic that gives the player god-mode control, overhead visibility, or special powers breaks the game's identity. The player is a citizen.
• Compute is always a constraint. Every AI feature must be evaluated against whether it can run on a normal PC using the priority/tier system. Grounded in reality at all times.
• Butterfly effect is sacred. The small initial randomness of Adam and Eve must be preserved. Do not add systems that flatten divergence or force worlds toward similar outcomes. The mutation system in language inheritance exists specifically to protect this principle across generations.
• Scarcity makes relationships matter. The Tier 1 cap of 8 is a design decision, not just a technical one. Players should feel the weight of who they invest in.

Open Problems (as of March 2026)

• Full implementation language / engine not yet decided
• Exact neural network architectures for emotion, senses, memory not finalized
• How social graph propagation works when player isn't present (NPC A teaches NPC B while player is across town) — current answer: unfamiliar NPCs don't get simulated deeply, so this interaction is shallow or deferred
• Economy, daily life systems, and world structure beyond the AI brain not yet designed
• Exact coherence layer architecture for Tier 2 grammar system not finalized
• How knowledge (farming, geography, history) is stored, whether it lives in the same weight structure as language or as a separate system
• Fast-forward catchup fidelity — how accurate does the simulation need to be vs how fast it needs to run during the loading screen