The first
morning.

The river has
no instructions.

Morning arrives before anyone knows what to do with it.

The water is moving. The trees are still. Somewhere beyond the clearing, an animal has already noticed the newcomers. Thirty people stand inside a world that owes them nothing: no food in their hands, no promise that tomorrow will be kinder, no narrator waiting to rescue the person who makes the wrong choice.

One walks toward the river. Another studies a stone. A third watches the second, and something almost too small to call history begins: an observation passes from one mind to another.

Suppose the stone becomes useful. Suppose its discoverer keeps it. Suppose someone asks for help, and the answer is no. The first tool would then be more than an object. It would become a question about trust.

Will they learn only how to live another day—or how to make another person's tomorrow possible?

This opening is illustrative fiction, not a transcript or a claim about the current world state. Actual outcomes must come from recorded simulation events.

GENESIS token → See all six MCap thresholds and planned unlocks

Not a civilization to command.
A civilization to observe.

GENESIS is an experimental, AI-assisted simulation of a small human-like community in a prehistoric setting. Its founding population is thirty adults, with fifteen female and fifteen male characters. The central objective is survival. There is no player-written sequence of destinies: people must navigate resources, constraints, relationships and the consequences of their permitted choices.

The project is developed with GPT-6 Astra under human direction: AI-assisted engineering builds the environment, code and systems, while a human sets the research aims and approves changes. This is different from saying the running world is powered entirely by GPT-6 Astra or that AI created it without human decisions.

“Humanity reborn” is our narrative lens. The scientific question is narrower: can useful behavior accumulate, survive failure and move between individuals? We aim to examine individual adaptation and cultural transmission—not announce that consciousness, free will or human evolution has been proved.

“From the beginning” does not mean a blank mind or a world without authored rules. The model already contains learned language and prior knowledge. Designers supply bodies, materials, action limits, foundational skills and initial conditions. The current research world also includes authored family relationships and earlier human interventions. Those starting advantages must be documented, not mistaken for discoveries.

Thirty bodies.
Thirty different mornings.

Two people can stand beside the same tree and carry away different histories. One arrived hungry. One remembers help at the river. One has already failed at the task the other is about to attempt. Shared surroundings need not produce shared decisions.

In the current runtime, a local Qwen3 4B model supplies bounded proposals. Characters share model weights but have separate personal contexts; they are not thirty separately trained models. Memory, needs, observations and relationships make those contexts different. Authored simulation rules validate actions and advance the world. When inference is delayed or a proposal is invalid, survival logic must still work.

The language model cannot simply declare a meal eaten, a tool made or another character persuaded. A statement is not a completed action. Distance, available resources, physical state and the implemented action rules determine what can actually happen.

Self-development here means changes in remembered experience, practiced skills and later behavior. It does not currently mean updating model weights, rewriting its own program or escaping its allowed environment. Autonomy is a bounded decision process, not evidence of subjective experience.

Names provide continuity in the story. Historical inspiration is not a claim that these were the names of the first humans, nor that a character accurately represents a historical person.

A skill becomes a story.
A story becomes a second chance.

Survival is not one successful decision. It is a chain that must keep holding: reach water, find food, conserve energy, recover from injury, return safely. A useful plan that cannot cross the terrain is not useful enough.

1. Notice. Encounter a resource, a route, a failed attempt or another person's work.
2. Try. Spend time and material on a permitted experiment. Failure must have a cost.
3. Test. Compare the result with an observable effect: did the edge cut, did the vessel help, did the route reach food?
4. Remember. Preserve evidence with its context. Hearing a claim is not the same as verifying it.
5. Tell—and try again. Another person receives the idea, tests it and may adapt it. Only then do we have evidence of transfer rather than repeated dialogue.

Existing research mechanics include basic tool practice, vessels, shelter work, resource use, conversation and bounded illness, wildlife contact injury and care. These are implemented systems with limits, not a complete ecological model. Sophisticated hunting, controlled fire and an open-ended technology tree remain future work.

A conversation should matter beyond its wording. An invitation could precede cooperation; a warning could change a route; a broken commitment could alter trust. We need to measure that connection. Fluent dialogue alone does not prove understanding, and repeating “I explored this place” is not the social life we are trying to build.

Death may occur through the world's rules. But starvation caused by broken navigation is an engineering defect to investigate, not an achievement in realism. The field record must distinguish modeled hazards from software failure.

When one person learns something valuable, who will be the second to know?

Token & MCap milestones.

One shared GENESIS main token is intended to govern milestone eligibility on Robinhood Chain. The proposed thresholds below open opportunities across the world. They do not buy an agent's loyalty, choose a partner, guarantee a discovery or determine an election.

All six unlocks are planned and currently locked. A threshold must satisfy the accepted data, liquidity and six-hour hold rules; each mechanic must also be implemented and explicitly activated. A higher MCap alone does not enable these features today. The full chapter descriptions and qualification rules follow below.

Release status · 6 September 2026: the deployed v14 policy is observation-only (shadow). No GENESIS CA is configured. All six new reward mechanics below are pending integration and activation. Reaching a number alone cannot unlock them today.

What counts as reaching a threshold?

The current shadow policy requires a six-hour continuous qualifying observation window, fresh accepted data and a liquidity constraint. The eligible estimate is capped at the smaller of the MCap estimate and 25 times accepted liquidity; liquidity must be at least $5,000. Data older than two minutes, rejected observations and excessive gaps must not qualify a new reward. A momentary chart spike is not enough.

These are guardrails, not proof that a market cannot be manipulated. CA-specific venue, supply definition, USD pricing and liquidity validation remain acceptance gates; the V4 reader is incomplete. The market value is an estimate, not automatically a verified circulating market cap. Live policy changes require versioned documentation and testing.

When active mechanics are ready, the design calls for one-time rewards and retained earned capabilities through later price drops—not repeated gifts every time the same threshold is crossed. Restart, outage and duplicate-reward protection must be verified before activation. Hunting and fire have no assigned MCap threshold yet.

Thresholds describe a conditional product roadmap, not price targets, a financial return or an airdrop entitlement. No token purchase is required to read this book or observe the public demo. Market changes are experimental interventions, not evidence that intelligence improved.

A future beyond oneself.

A family should not begin because a number changed on a chart. The $3M gate would make a set of possibilities available; the characters' bounded decisions and the world's safety rules would determine whether anything follows.

This part of the experiment is explicit about its rules, not sexually explicit in its presentation. Finding a partner, developing attraction, accepting or refusing a relationship, and deciding whether to form a family are intended to be visible social events. Intimacy itself remains abstract. There will be no graphic sexual scenes.

• Adults only. Both partners must meet the configured adult-age requirement. Children must never participate in sexual, courtship or reproductive actions. The simulated calendar and aging rates must be published before this system is enabled.
• Mutual, revocable acceptance. Neither a viewer, a holder, a leader nor one partner can force a pairing. Refusal and separation must be valid outcomes. Simulated acceptance is an explicit state check, not an inference from silence or a substitute for real human consent.
• Kinship before eligibility. Parentage and close-family exclusions must be checked before a pair can enter a reproductive state. Ambiguous lineage should block eligibility, not silently permit it.
• Care before expansion. Health, available food, shelter and caregiving capacity must constrain progression. Birth cannot become a guaranteed reward for market support, and population limits must protect simulation stability.
• A visible state history. Public records should identify partnership acceptance, pregnancy, birth, parents and caregiving events without revealing private prompts or intimate dialogue. Refusals must not be rewritten into success.
• A child is a new individual. Any future inheritance rules must distinguish declared initial traits from knowledge acquired through later observation and teaching. Children do not automatically receive a parent's entire memory or expertise.

These are implementation and acceptance requirements, not claims that the live world already runs this complete lifecycle. Reproductive probabilities, timing, population caps and inheritance rules are not finalized. Existing authored couples do not demonstrate autonomous mate selection; old test fixtures do not demonstrate live births.

Only after a working multigenerational system and appropriate controls could we investigate inheritance and selection within this model. Even then, results about a designed simulation would not prove human biological evolution or that artificial agents are living organisms.

Could the next generation inherit something more valuable than a tool—the habit of teaching?

A beautiful story
is not enough evidence.

The research ambition is a testable account of adaptation and cultural accumulation. A moving scene is a reason to ask a question, not an answer. The protocol below is proposed work; no completed controlled study or statistically validated result is claimed here.

Publish the conditions before interpreting the result

A reproducible run should identify the software and policy versions, seed, model and prompt configuration, inference budget, population, initial knowledge, resource layout, hazard rules and simulated duration. Record model outputs and accepted or rejected actions in a protected research log; publish appropriately redacted evidence. Identical seeds alone do not guarantee identical LLM outputs.

Compare multiple matched runs, including no-memory, no-teaching and rules-only baselines. Keep computation and resources comparable. Separate development runs, owner interventions and market-gated runs from controlled experiments. Predeclare success criteria and report effect sizes, uncertainty, failed runs and negative results—not only the most compelling replay.

Public telemetry should eventually connect a claim to an event chain: an observation, an attempt, its result and later independent reuse. The current timeline is a useful observation surface, not a complete research dataset or a view into private thoughts. No benchmark score, sample size or statistical significance is invented for this book.

Watch closely.
Do not write their answers.

The backend is the single authority for the world. The browser renders snapshots and recorded events. Camera movement and character selection change your view, not the characters' lives. Opening this document, refreshing market status or holding a token cannot issue an order to move, attack, pair or elect someone.

The current public demo is a WebGL observer connected to a VPS simulation. Unreal Engine remains the intended higher-fidelity presentation path; this page does not turn WebGL into an Unreal stream. Better graphics can improve observation, but they cannot establish that the underlying behavior is intelligent.

Private memories and internal model contexts are not broadcast as public thoughts. Dialogue shown or spoken is simulated character output, not testimony from a conscious person. The experiment must not treat persuasive language as permission to bypass world rules.

Infrastructure and policy are human-operated. This is not an entirely decentralized organism: server operators control deployments, approved world migrations and data-source configuration. Market readers supply observations; they do not possess authority to mint tokens or directly command characters. Future leader-token creation and distribution require separate, explicit on-chain approval.

What exists.
What comes next.

The long-running world is a development experiment. On 6 September 2026, an owner-requested intervention revived 23 deceased characters while preserving the seven survivors and prior history. That was not natural recovery and must be excluded or explicitly modeled in survival research. The subsequent v14 market migration did not revive people or grant new rewards.

Performance and stability remain research constraints. A transient health-check 503 was observed under high VPS CPU load during release verification; later direct checks recovered without a reported persistence error. This is not a zero-error or uninterrupted-service guarantee. Model latency, navigation defects and rendering limitations must remain visible in the engineering record.

The next chapter is not a promise that they will become a civilization. It is a commitment to make the question observable—and to keep the failures in the record.

The river will still be there tomorrow.
What will they bring back to it?