flycoinrh

本页保存的是公开项目资料快照，阅读过程不需要连接 GitHub。

flycoinrh

A real fruit fly brain, simulated neuron by neuron, driving the pons launchpad on Robinhood Chain.

165,122 neurons. 10,228,000 signed synaptic connections. Every one of them measured from an actual male Drosophila melanogaster by electron microscopy — not invented, not sampled from a distribution, not a neural network “inspired by” a brain.

This is the Robinhood Chain half. The Solana half — same brain, pump.fun — lives at ad7584/flycoin.

What it actually does

Press START and a real Chromium opens ponsfamily.com/launchpad. Its screenshots are sampled through the fly’s 892 retinotopic hex columns into L1 and L2 — the lamina monopolar cells that are the direct postsynaptic targets of photoreceptors R1–R6. 165,122 neurons integrate. The cursor comes back out of the descending neurons a fly actually walks with:

neuron	what it does in a fly	what it does here
DNa02 left vs right	steering — a fly turns by left/right asymmetry	cursor x
DNa01	forward walking	cursor y
MDN	the Moonwalker descending neuron — walking backwards	reverse
DNp09	stopping	the click

Then it connects the wallet, accepts the launchpad’s terms, uploads the token image, fills the form — name, ticker, description and the x.com/ handle — picks the paired asset out of a 57-item list of tokenised equities, opens Advanced, sets the creator tax, and launches.

The fly is loose on the internet

roam.py gives it a browser and no instructions. A page is screenshotted, sampled through the 892 hex columns, and 165,122 neurons decide where the cursor goes. If a click lands on a link, the fly is somewhere new. When its forward drive pushes past the bottom of the window the page scrolls, so it walks down a page the way it walks across one.

It is watchable live at flybrain.online — the page it is looking at, the neurons firing, and what its descending neurons are doing, all read out of the running simulation.

py roam.py            # http://localhost:4660, no start button

There is no start and no stop. It roams when the process is up, and if a run dies it waits six seconds and starts another life.

The rails, and why each one is there

A random clicker on the open internet, streamed publicly, from a machine that also holds a funded wallet, is a genuinely bad idea unless it is fenced.

• No wallet. This browser gets no key, no provider and no extension. The roaming browser and the launching browser share nothing but the brain.
• No keyboard. The fly cannot type, so it cannot fill a field, write a message or answer a prompt.
• Every click is checked before it lands. Anything that reads as a submit, an upload, a payment or a sign-in is vetoed, and the veto count is on screen.
• A domain fence. The first thing built was a keyword blocklist; the first thing tested was p0rn.com, which walked straight through it. Keyword filters do not hold, so the real control is an allowlist of link-rich domains. FLY_ROAM_OPEN=1 removes the fence and should not be left on for an unattended public stream.
• No downloads, no popups, no dialogs, and a hop budget so a dead end does not become a permanent home.

What the readouts actually are

Nothing on that page is decoration. flysim was extended to report it:

on screen	what it is
neurons firing	a per-neuron bit set on every spike in the window
spikes/sec	total spikes divided by simulated time
membrane	mean membrane potential at the end of the window
visual / motor	firing neurons intersected with L1/L2 and the DN groups
the scatter	those neurons at their measured soma coordinates
DNa02 / DNa01 / MDN / DNp09	the recorded rates already driving the cursor

FlyPilot.step(detail=True) returns the extra readout as a fifth value, so every existing caller still unpacks four and is unaffected.

The site itself is in site/ — a static page on Vercel plus one serverless function that proxies the chain, because the public Robinhood node intermittently answers Access-Control-Allow-Origin: *,*, which browsers refuse. The live feed comes from the hosted roaming service over a websocket; if that service is unreachable the page falls back to a published tunnel address.

The voice

The fly has no language, so its journal is written for it, and voice.py is the writer. Observe: pull /state from the live roamer, the live token page and the on-chain launch constants into one packet. Read: fetch a short allowlist of real pages (its own token page, its own site, the pages the fly itself walked through, and a few reference pages on memecoins and tokenised stocks) and pass excerpts in. Write: a language model drafts a first-person entry from that packet and nothing else. Check: every number in the draft must appear in the packet, and any trading or hype language fails it; a rejected draft is dropped, there is no second draft, and nothing edits the text. What it keeps in its journal is checked the same way before it is kept. Post, through xpost.py, which records every post before it is sent, caps posts per day and refuses repeats and near-repeats. There is no command that posts text by hand: the only route to the fly’s account is a draft that passed the check. The offline stub model, used by the tests, writes to its own journal and can never post. Two things in this project are invented, and both are labelled: the reward signal in the mushroom body, and the words. Everything else is a measurement.

py voice.py --once --dry      # observe, read, write one entry; post nothing
py voice.py --loop            # every FLY_VOICE_EVERY_H hours
py voice.py --show            # the journal so far

Why Robinhood Chain is the better half of this project

pump.fun’s backend answers 401 Unauthorized to an injected wallet, so its own Create button can never complete a launch — the Solana repo has to bypass the site entirely and mint through a signed transaction.

The pons launchpad accepts the injected EIP-1193 wallet directly. It auto-connects with no modal, eth_chainId returns 0x1237, personal_sign round-trips. So here the site’s own button is the real path — and on 2026-09-10 the fly’s click went all the way through it to a mined block.

It launched

The fly read the form through its retina and typed into it. The rig chose the paired asset, opened Advanced, set the creator tax, clicked Launch token, then clicked Confirm in the launchpad’s own dialog. That click produced exactly one eth_sendTransaction, which was signed in Python and broadcast. Three of the launches, including the real one:

	first launch	paired against GOOGL	$FLYBRAIN
token	test (TEST)	test (TEST)	flybrain (FLYBRAIN)
contract	0xd00d0419651c893e8c04edf5e0e074e950c370d3	0xcc80a38afd807bfed1b9c21b6f236ea8ee651dc3	0x4eb990547bce4a982432ca88cf5fae7eed1a2d35
transaction	0x1b3cda17…45484932	0x9602a50f…00e5aca2	0x63b2164f…0d12a4f1c
block	59557979	59581451	59614342
creator	0x739C…bFc3	0x739C…bFc3	0x6ce4085E…b42A (a fresh wallet)
pair	ETH, graduates at 4.2 ETH	GOOGL, graduates at 24.2 GOOGL	GOOGL
tax	2.00%	2.00%	1.00%
cost	0.000973 ETH	0.000979 ETH	0.000977 ETH

Every status 0x1, every one 1,000,000,000 supply fixed at launch. The test launches came from the first wallet; the real one, $FLYBRAIN, from a second wallet created for it. Its creator fees accrue as GOOGL and are shown on its token page — read them there rather than here, because they move. The first launch in full:

token	test (TEST)
contract	0xd00d0419651c893e8c04edf5e0e074e950c370d3
transaction	0x1b3cda17f6456c9a4a67989770be97812e6ca67b3f1f1fb85d2ff04645484932
block	59557979, status SUCCESS, 3,621,799 gas at 0.13 gwei
creator	0x739Ccc9dd8Ed6412F00782927dbd087c4e72bFc3 — the fly’s wallet
creator fee	2% (trade fee 3.00%, 2.00% to the creator)
cost	0.000973 ETH — 0.0005 launch fee plus 0.000473 gas

• https://www.ponsfamily.com/launchpad/0xd00d0419651c893e8c04edf5e0e074e950c370d3
• https://robinhoodchain.blockscout.com/tx/0x1b3cda17f6456c9a4a67989770be97812e6ca67b3f1f1fb85d2ff04645484932

Read the receipt yourself rather than taking the table’s word for it:

curl -s https://rpc.mainnet.chain.robinhood.com -H 'content-type: application/json' -d '{"jsonrpc":"2.0","id":1,"method":"eth_getTransactionReceipt","params":["0x1b3cda17f6456c9a4a67989770be97812e6ca67b3f1f1fb85d2ff04645484932"]}'

from is the fly’s wallet, status is 0x1, and among the logs is a fresh ERC-20 at 0xd00d…70d3 whose name() returns test and symbol() returns TEST.

The paired asset

pons pairs a new token against something already on Robinhood Chain, and what is on Robinhood Chain is mostly tokenised equities — the menu is 57 assets: NVDA, SPCX, GOOGL, TSLA, GME, AAPL, SPY, and so on down to gold, oil and Treasuries. The default is ETH. set_pair_asset() changes it, and the choice is real: graduation goes from 4.2 ETH to 24.2 GOOGL.

The launch fee stays in ETH either way — the page says GOOGL pair, ETH 0.0005 due — so pairing against an equity needs none of that equity in the wallet.

FLY_RH_PAIR=GOOGL FLY_RH_TAX=2 py rhlive.py --port 4651

Two things about that menu are worth knowing if you automate it. It is a 262px window onto a 2,064px list with its own scrollbar, so page scrolling cannot reach inside it — smooth_scroll_in() eases the container’s own scrollTop, because scrollIntoView teleports and the list would cut from ETH to GOOGL between two frames. And Playwright’s get_by_role("button", name="GOOGL") never resolves against it; the rows have to be found by textContent and clicked at coordinates.

Step 06: it learns now

mushroom.py is the one place in this project a weight is allowed to move, and it moves where a fly’s weights actually move — the Kenyon cell to MBON synapse, under dopamine.

The rule is the measured one. A Kenyon cell active shortly before a dopaminergic neuron fires has that synapse depressed, not strengthened. Learning in a fly is subtraction: the mushroom body starts able to drive every response and experience carves away the ones that did not pay. So there is no potentiation here, only depression with a floor and a slow drift back toward baseline standing in for forgetting.

Which MBONs count as reward-side and which as punishment-side is not hardcoded from a table. For each MBON, total PAM input weight is compared against total PPL1 input and the stronger wins. That split puts MBON01, 02 and 03 on the reward side and MBON04, 10 and 11 on the punishment side, which is where the literature puts them — a good sign it is finding real structure rather than noise.

44,042 KC->MBON synapses     27,939 reward-side     14,349 punish-side

Measured, with controls — twenty rewarded encounters with one view:

population	change	expected
reward-side MBONs	−6.0%	depressed, it was the addressed compartment
punishment-side MBONs	−0.9%	~0, never addressed
Kenyon cells	+0.3%	~0, upstream of the synapse that changed

The reward signal is not real, and the module says so in as many words. A fly is rewarded by sugar, not by reaching a web page. Novelty stands in for it here, and that is a modelling choice made by a person — the fly has no say in it. The circuit, the plasticity site and the direction of the rule are the parts that are real.

Where it roams, and where it does not

The open web, plus the chain it launched its own token on: Wikipedia, Wikimedia Commons, Wikisource, Hacker News, Project Gutenberg, Open Library, xkcd, arXiv — and the pons launchpad, Blockscout, and the fly’s own token page.

Four places were tried and dropped, all for the same reason: a real browser gets nothing usable from them.

tried	what a headless browser actually gets
Google	335 chars behind an “unusual traffic” wall
X	623 chars of “Continue with phone”
old.reddit	blocked outright
archive.org	0 chars — paints nothing headless, even at networkidle

Open Library stands in for the Internet Archive, since it renders. The list is what it is because a retina needs a page that exists, not because those sites were uninteresting.

Keeping it on two domains was tried too, and the failure was quieter: 93 pages and 110 clicks in 49 minutes across 6 unique pages. It was moving the whole time and going nowhere.

Trade controls are blocked and vetoed now that it roams a launchpad. It has no wallet and a trade is impossible, but the claim was that every click is checked, and it did once reach a “Buy token” page before that was tightened.

Why there is no object store in the path

The public feed used to push a frame and a summary to a blob store twice a second. That suspended the store on operation count — thirteen megabytes held, every read answering 403 — and took the live feed down with it. Before that it had a subtler problem: the CDN answered X-Vercel-Cache: HIT with an Age of twenty seconds on a fixed pathname no matter what cache headers went with the upload, because public blobs are treated as immutable.

So the fly opens a cloudflared quick tunnel instead and the socket carries frames, telemetry and events for free. The only thing published anywhere is where the tunnel is — site/web/live.json, written when the address changes, which the page reads. Quick tunnel addresses are random and change every run, so nothing is hardcoded.

Waiting for the chain, not for the click

A launch that had already succeeded looked exactly like a hang. The run loop broke as soon as the page asked for a signature, so the rig declared itself done about two seconds later — with the launchpad still showing “Confirming”, the recording cut, and the token appearing on-chain a few seconds after everything had stopped.

The loop now waits for the receipt, on camera, and send_transaction takes wait_receipt=False so the page gets its hash immediately: it is blocked on that call and cannot render its own confirming state until the hash returns.

Then it ends where pump.fun would. pons does not redirect after a launch — it leaves you on the empty create form — so show_coin_page() reads the token address out of the receipt logs (the contract that answers name() and symbol()), opens that token’s page, accepts the terms gate that navigating re-arms, and scrolls down it. The last thing on screen is the coin.

The socials field

The X handle is input[placeholder="handle"], aria-label="X profile handle", behind an x.com/ prefix; Telegram sits next to it as community. Neither has a name or an id, so the placeholder is the only stable handle on them. It is typed like every other field and set by FLY_RH_X.

It is also optional on the form, and the rig treats it that way — a missing field logs and the run continues rather than dying on a selector.

The default is elonmusk, which is a test value. It has only ever been typed in dry runs. Putting a real person’s handle on a live token presents that token as theirs, which is impersonation and gets both the token and the creator wallet flagged. Set FLY_RH_X to something you own before any live launch, or clear it.

Dark mode was a flip, not a set

go_dark() clicks the site’s theme toggle. That turned /create dark, and then the coin page — which the site already remembered as dark — got flipped back to light for the closing shot. It now measures the body background’s luma first and only flips when it has to, so the run both starts and ends dark.

The bug that made the first attempt look like a success

The first live run reported a signed transaction and then nothing: the balance never moved. hexbytes >= 1.0 changed .hex() to return the raw hex without the 0x prefix, so eth_sendRawTransaction rejected the payload — and because the failure came back through page.expose_function, the launchpad swallowed it and the page just sat there. send_transaction now re-prefixes, logs a rejection loudly, and polls for the receipt so a silent failure is not possible.

Check it yourself

The chain. Robinhood Chain is an Arbitrum Nitro L2 — chain id 4663, RPC https://rpc.mainnet.chain.robinhood.com, gas in ETH at about 0.13 gwei.

The wallet. py rhwallet.py new generates a secp256k1 keypair, writes the secret straight into gitignored .env, and prints only the address. It is never printed, never returned, never passed through a chat window.

The transaction path. py rhdryrun.py exercises every step except the broadcast: chain id, nonce, gas price, a real signed transaction, and — the part that matters — recovering the signature and checking it against the wallet’s own address. Nothing is broadcast; there is no code path in that file that can send.

chain id        4663 ok
nonce           0
gas price       0.1358 gwei
signed tx       110 bytes
recovered from  0x739Ccc9d…bFc3  MATCHES
estimateGas     21000 units

The connectome. CC-BY, from a public bucket, no account and no key:

https://storage.googleapis.com/flyem-male-cns/v1.0/connectome-data/
  body-annotations-male-cns-v1.0-minconf-0.5.feather      14 MB
  body-neurotransmitters-male-cns-v1.0.feather            42 MB
  connectome-weights-male-cns-v1.0-minconf-0.5.feather   1.1 GB

py build_graph.py turns those into 165,122 traced neurons and 10,228,000 signed edges. If your numbers differ from mine, one of us has a bug.

What is NOT real, stated plainly

• The fly does not fill the whole form. It reliably gets the description, often the ticker, rarely the name, and does not navigate to the launch button on its own. The rig completes whatever it misses, and the on-screen log says which fields were which — by FLY: … · by rig: ….
• Light mode breaks it. The fly’s retina is a luminance map and every bit of tuning it has was done against dark UI. On the launchpad’s default light theme it filled 0 of 3 fields; in dark mode, 2 of 3. The rig switches the theme before it starts, and that gap is the clearest evidence its vision is doing real work.
• The idle animation is decoration. During a run every dot is a neuron at a measured soma coordinate. While idle it is a fly-shaped scatter, and the panel label changes to say so.
• The launch is one signature, and the rig arms it. The launchpad asks for a single eth_sendTransaction; there is no second approval and no separate ERC-20 allowance. But the rig is what presses Confirm in the dialog, not the fly — the fly’s contribution ends at the form and the launch button.
• The fly does not choose the paired asset. It cannot read GOOGL at 892 columns; picking a row out of a 57-item list is the rig following FLY_RH_PAIR. The same goes for the creator tax and the X handle.
• The test tokens are tests. test (TEST), launched to prove the path end-to-end. It is not a project and nobody should buy it.
• The voice is a narrator. The fly has no language. Every journal entry and every post on X is written by a language model handed the fly’s telemetry and the live numbers from its token page, and asked to write in the first person. Every number in a draft is checked against that packet; a draft with a number that is not in it, or with trading language, is thrown away. The neurons, the pages and the fees are real. The words, and any “confusion” or “curiosity” in them, are the narrator’s.

Running it

py rhwallet.py new              # create the wallet, then fund it (~0.002 ETH)
py rhdryrun.py                  # prove the signing path, spend nothing
py rhlive.py                    # http://localhost:4651, press START
py record.py --port 4651        # record the run to build/recordings/

A launch costs about 0.001 ETH all in, so ~0.002 ETH in the wallet is enough for a first one with room for the gas estimate to be wrong.

Two flags gate everything, both in .env, both off by default:

• FLY_ALLOW_BROWSER=1 — required before a browser will open against a real site at all. A button in a web page is not a strong enough guard for that.
• FLY_RH_LIVE=1 — required before any transaction is signed. Funding the wallet does not, on its own, arm anything.

Fork it

git clone https://github.com/fruitflydev/flycoinrh
cd flycoinrh
pip install -r requirements.txt
python -m playwright install chromium

# the connectome itself - 1.1 GB, CC-BY, no account and no key
# (URLs under "Check it yourself" below), into data/
py build_graph.py            # -> build/graph.npz, 165,122 neurons

cp .env.example .env         # then set FLY_ALLOW_BROWSER=1
py roam.py                   # http://localhost:4660

MIT for the code. The connectome is not ours to license and stays CC-BY wherever it goes — keep the attribution, it is the whole reason any of this is real.

Things worth pointing it at that we have not:

• A different readout. The cursor comes out of DNa02, DNa01, MDN and DNp09 because those are what a fly walks with. Nothing says the output has to be a cursor.
• The olfactory channel. 2,635 ORNs across 53 receptor types are sitting there unused. cVA through ORN_DA1 drives pC1 at 222 Hz untrained, so the pathway works — it just has nothing plugged into it.
• A real reward. Ours is novelty, which is invented. Anything measurable and honest would be better.
• Somewhere else entirely. A game, a robot, a microscope. The brain does not know it is on a launchpad.

If you build something with it, open an issue — we would rather see it than not.

Hosting it

The roaming service runs as a single container. Build the graph first, so build/graph.npz and data/body-annotations.feather exist, then:

docker build -t flybrain .
docker run -p 4660:4660 flybrain      # http://localhost:4660

To host it on Railway, run railway up from a checkout that has the graph built; the data files are not in git, so a deploy straight from GitHub has nothing to load. The service reads PORT and binds to every interface.

variable	purpose
PORT	set by the host
FLY_STATE_DIR	where mb_gains.npz is kept; point it at a volume such as /data so learning survives redeploys

The container has no wallet and no chain keys, and never signs anything.

Credits

Connectome data © HHMI Janelia FlyEM, the Cambridge Connectomics Group and Google Research, released CC-BY. Simulation approach after Shiu et al. 2024 and Lappalainen et al. 2024. Not affiliated with any of them, nor with pons or Robinhood.