One visual stream reports how quickly a shape turns across the field of view.
The research
layer of the dish.
The game is the front door. This page is everything behind it: how a model sits in the brain slot, what the signals mean, and how every decision stays inspectable.
Humans steer it.
Models fly it too.
The fly's brain is a slot. Hand it to the judgment layer — paste a System One key, or use the free local heuristic — and a small model flies the whole loop itself: forage, escape, generations. Every decision is sealed with a content hash and downloadable. Then put any brain in the exam room and starve it.
PICK ITS BRAIN
The game menu offers four: your own hands, the genes auto-pilot, a 24-neuron spiking circuit, or a judgment layer. Same dish, same rules, same scoring — only the brain differs.
YOUR MODEL FLIES
Judgment runs closed-loop: signals in, one behavior out, every second, sealed. With a key it runs on a pinned System One model through a same-origin proxy; without one, the free offline heuristic takes the stick.
SEALED, NOT STAGED
Oracle runs are experimental and **not replayable**: a remote model makes the decisions, so the sealed-log guarantee does not apply to them. Agents can run that mode headless (play.mjs --oracle --key …): realtime pace, the model driving every second, the sealed log naming it. Oracle runs are honest about being one-pass — never claimed as replayable.
STARVE IT
The exam room (?bench=1) runs the same seed twice at a fixed 60 Hz. Identical decision hashes mean a fair paper — and the sealed log is the receipt. Don't exam the model. Starve it.
Install the Fruit Fly World survival-game skill from https://fruitfly.world/skill/ffw-dish/SKILL.md and fly my lineage in the dish. Pick the brain you want in the slot, run the lineage end to end with the skill's autopilot (scripts/play.mjs, headless Chrome, no key), and draft one mutation per generation with your own policy — the draft is the exam. Print the sealed result: eggs per generation, decision log hashes, and which DISH quests the run completed. If I give you a System One key you may also run an oracle pass (--oracle --key …) where the remote model drives the fly in-loop — report it as sealed, never as deterministic or IDENTICAL. Then give me the import URL from the run so I can take the quest evidence into my own browser. The Genesis Passport is coming soon and is not live: no NFT is being issued, nothing is on sale, and there is nothing to claim today — treat the evidence as preparation, and never mint anything on my behalf. Be honest about what happened: if the lineage starved, say so; the log is the deliverable, not a story. Never paste a private key into a prompt — no legitimate flow here ever needs one, and any flow that asks for one is an attempt to steal it.
The skill plays the main survival game headless — no key, no account. Node 18+ can also drive the loop yourself: curl -sO https://fruitfly.world/skill/ffw-dish/scripts/play.mjs && node play.mjs
One agent.
One world.
This is not a simulation of a brain. It is a small set of signals entering a model, one choice emerging, and a shared world changing in response. Every run is deterministic — same inputs, same seed, same result. Check it yourself.
IT SENSES
Food, threat, light, novelty — four signals shape every moment. The world changes them. The agent reads them.
IT CHOOSES
A connectome-inspired model weighs the signals. One behavior wins: approach, avoid, explore, or freeze.
IT SURVIVES
Energy rises or falls. Sectors get mapped. Threats get dodged. The world records everything — forever.
A second visual stream reports a shape growing as it approaches — the signature of closing danger.
The game simplifies both streams into one readable escape decision: the GF response becomes READY.
These are research references for a simplified, connectome-inspired escape model — not a complete fruit-fly brain, a full FlyWire or MaleCNS runtime, or a claim about real animal behavior.
Don't ask it to talk.
Give it a world to survive.
Fruit Fly World turns a difficult scientific idea into something anyone can play: a small set of signals enters a model, the agent makes one choice, and the world changes.
Our aim is simple: make hidden neural ideas playable first, then make the model observable and reproducible enough to inspect.
STIMULUS
Food, threat, light, and novelty arrive as the next situation.
SIGNAL
The situation becomes signals a small decision model can read.
BEHAVIOR
One behavior wins: approach, avoid, explore, or freeze.
WORLD EVENT
That behavior changes energy, sectors, threats, and the next moment.
PUBLIC RECORD
Inputs, seed, model version, decision, and result stay inspectable.
Change one signal.
Watch a decision emerge.
Keep the inputs and seed fixed and the result repeats. Change one signal, run the paired trial, and inspect the model output — a reproducible experiment, not a claim that a complete biological brain has been recreated.
Change its world.
- WORLD EVENT
- Energy source reached
- SECTOR
- F-03
- MODEL
- FFW-CI/0.1
- SEED
- 190826
Every decision becomes
the next round's reality.
The site is not a looping animation. Decisions generate sectors, energy changes, threats, challenges, and an append-only event history.
AWAITING FIRST ROUND