Can a fly smell a rug pull?

Research finding, not financial advice

A real fruit fly connectome (165,122 neurons, the same one RHABDOM runs) was shown real Pons launches on Robinhood Chain and trained to associate a launch's smell with what actually happened to it on chain. Then it was tested on launches from a later time window it had never seen, and measured against a coin flip, a fixed checklist, a plain logistic regression, and its own scrambled-wiring twin.

What this is not. Nothing on this page is a signal to buy, sell, or avoid any token. It is a measurement of whether a simulated insect brain, trained on a few hundred past launches, does better than chance at a narrow, historical question. No wallet, no key and no trade exist anywhere in this codebase; the chain connection only reads.

A conflict worth stating. RHABDOM has its own token on Robinhood Chain (0x9df2066E21233Ad40F3657b632c726d5B7eedd2F). That is a reason to trust this page less, not more. The numbers below were measured before it existed, they are published whether or not they flatter the method, and the method caught 0 of 14 real dumps — which includes having nothing useful to say about our own token.

The question that turned out to have an answer

Three different "does a dump come later" questions were tried on the exact same 350 training-window and 180 test-window launches, in order, and every one is reported here — not just the one that worked.

Look at…→predict by…	Look	Horizon	Min buyers	Train eligible	Train rugged	Test eligible	Test rugged
1 hour → rest of day	36,000	864,000	5	103	1	61	0
5 minutes → rest of hour	3,000	36,000	2	140	2	89	0
10 seconds → next 5 minutes (used below)	100	3,000	0	329	52	166	22

Blocks, not minutes, are what the protocol is fixed in; the labels above are the nearest round number at this chain's measured ~0.1015 s per block. The first two rows found almost no dumps happening after their look — on Pons, a creator dump is front-loaded, not spread over the day. That is a real, disclosed finding, and it is why the task below looks only 10 seconds after launch.

The task actually tested

At 10 seconds after a token launches — before most outside buyers have even arrived — will the creator's wallet (and anything it has passed tokens to) dump at least half of its peak holding, for at least half of everything outsiders have paid in, within the next 5 minutes? Training used 150 eligible launches from 12–20 days before the frozen head block; testing used 90 from a separate, later 1.1–1.5 day window. No launch and no time period is shared between them.

Base rate: 15.6% of the test launches were dumped within 5 minutes. A rule that always guesses “safe” would already be right 84.4% of the time — that is the number every result below has to beat, not 50%.

Check a Pons token yourself

Paste any token contract address that launched on Pons (Robinhood Chain). This reconstructs what was visible about 10 seconds after that token's own launch and runs it through the same three predictors below — live, on chain, right now. Nothing here watches for new launches or runs unless you ask it to; it reads the chain and answers once.

Read the result next to the bottom line below. This tool caught 0 of 14 real dumps in testing. A “safe” result here is not a safety signal — it is what a method with a 0% catch rate says about a token it has never scored before.

Reads Robinhood Chain over the public RPC; one check per address every few seconds.

What the fly smelled

Eight on-chain facts, each wired to its own antennal-lobe glomerulus in the real connectome (a fixed, disjoint assignment, not a claim that a fly would recognise any of this as a real smell):

Feature	What it measures
creator_launch_share	Share of supply the creator bought in the launch block itself
circle_share_at_look	Share of supply the creator's wallet (and anything it fed tokens to) holds at 10s
top5_share_at_look	Share of supply the five largest holders hold at 10s
snipers	Outside wallets that bought within ~3 seconds of launch
outside_buyers_at_look	Outside wallets that had bought by 10s
sell_pressure_at_look	Sale proceeds ÷ purchases so far, in the quote asset
circle_sold_at_look	Share of its own peak holding the creator's circle had already sold by 10s
prior_launches	Tokens this same deployer had already launched

At only 10 seconds in, most of these are still near zero for most tokens; the two that actually carry signal this early are the creator's own launch share and its snipers.

Result

Bottom line. The real connectome caught none of the 14 launches that were actually dumped in the test set — it scored close to the “always guess safe” baseline (83.3% vs 84.4%) by predicting almost everything as safe. Its own scrambled-wiring twin caught more of them (3 of 14) with the identical training and the identical eight channels — evidence that whatever small signal exists here is not coming from the real wiring. The fixed checklist did worse than doing nothing (55.6% accuracy, 28 false alarms on 76 safe launches), which is its own honest result: a simple rule written in advance, without seeing this data, was actively misleading on it. The plain logistic regression, fit directly on the same eight numbers, is the only predictor here that beat the majority baseline on accuracy (85.6%), and it still caught only 2 of 14. Read plainly: on this test, nothing here reliably tells a coming dump from a safe launch 10 seconds after it starts. That is the answer this ran to find, not the one it was built to prove.

Predictor	Accuracy	p vs. coin flip	Caught / recall	Precision	Mean score, rugged	Mean score, safe
Fly — real connectome	0.833	0.0000	0.0%	0.0%	0.238	0.222	44,042 KC→MBON synapses, 37,538 depressed by training
Fly — twin (scrambled wiring)	0.811	0.0000	21.4%	33.3%	0.020	0.023	1,019 KC→MBON synapses, 973 depressed by training
Logistic regression (fit on train)	0.856	0.0000	14.3%	66.7%	0.252	0.387	8 features, standardised
Fixed checklist	0.556	0.3428	14.3%	6.7%	0.714	0.263	launch share ≥ 15% or snipers ≥ 3

Accuracy alone is misleading here: always guessing “safe” already scores 84.4%. Recall on the rugged class — of the launches that actually got dumped, how many did each method catch — is the number that matters. p is an exact two-sided binomial test of the predictor's raw accuracy against a coin flip; it says nothing about the accuracy-vs-majority-baseline comparison above it.

What the logistic regression learned

Fit on the training split only, on standardised features. A positive weight pushes toward predicting a dump.

Feature	Weight
creator_launch_share	+0.066
circle_share_at_look	+0.909
top5_share_at_look	-0.281
snipers	-0.499
outside_buyers_at_look	+0.107
sell_pressure_at_look	-0.537
circle_sold_at_look	+0.261
prior_launches	-0.795

What is measured, what is chosen, what is not claimed

Measured

Chain: Robinhood Chain, 4663
Launchpad: Pons v2, factory logs
Train launches sampled: 350
Test launches sampled: 180
Train used (compute budget): 150
Test used (compute budget): 90
Kenyon cells firing per look: ~1257 of 4,064
Look length: 200 steps (40 ms)
Training epochs: 1

Chosen by people

The dump rule: ≥50% of the circle's peak holding sold, for ≥50% of what outsiders paid in.Same rule, unchanged, across all three time windows tried.
The smell wiring: eight features, one glomerulus each, value 0–1 of a fixed scale.A disjoint channel per fact, not an attempt at a real fly odour.
Kenyon-cell sparsity: PN ×0.5, APL ×10, KC ×0.3.Reused from flycoinrh's own measurement on this connectome, not refit here.
The threshold: leaning ≥ 0 predicts safe.Fixed before this ever touched real data.
The checklist: launch share ≥ 15% or snipers ≥ 3.Written before training, never fit to this data.

Not claimed

Advice. This is a historical measurement, not a live signal. Nothing here watches new launches or tells anyone what to buy.
A large sample. 150 training and 90 test launches is enough to see whether a signal exists at all, not enough to certify one for use.
That the twin proves nothing. Same neurons, same eight channels, same training — only the wiring differs. If the real brain and its twin score alike, the wiring is not doing the work.
That the earlier, degenerate tasks were failures. Finding that post-launch dumps are front-loaded is itself a real result about how Pons launches behave.

Connectome: male CNS v1.0, © HHMI Janelia FlyEM, the Cambridge Connectomics Group and Google Research, CC-BY 4.0. Chain data: Robinhood Chain public RPC, read-only. Part of RHABDOM.