We dropped a simulated fruit fly brain onto a randomly chosen criminology paper and let it walk outward through reference lists. Open-access papers are corridors. Paywalled papers are walls. It explores until it runs out of things it can read.
An illustration of the idea, not the simulation. The measured run is below.
On 3 September 2026, about a hundred scientists at HHMI Janelia, the University of Cambridge, the MRC Laboratory of Molecular Biology and Google Research published the first complete connectome — a full wiring diagram of a brain, every cell and every connection between them — of an adult male Drosophila central nervous system in Cell: 166,700 neurons and 11,710 neuron types, every one of them traced. Google posted about it and it went round the internet. The data itself, MaleCNS v1.0, had been out since June under an open licence, so anyone could download a whole fly brain and run it.
People did, immediately. Within a week the fly was playing Doom and Super Mario 64, walking around Minecraft, and driving a car in a simulator. Each of those works the same way: give the fly's neurons something to sense, let them spike, and let their output pick the next move. We pointed the same trick at something we care about.
Matthew Ashby checked 12,541 articles across 97 criminology journals published 2017–2019 and found that 22% were available to non-subscribers — that is, readable by someone without a paid journal subscription or a university library that has one. More than three quarters of the field was hidden.
A percentage like that counts what is open. It does not tell you what happens to a person doing research, because nobody reads a percentage: you start at one paper, look at what it cites, read what you can get, and follow that outward. Whether the field is usable depends on how far that chain gets before it dead-ends. We wanted to watch that happen instead of assuming it, so we built a reader we could actually run and count: the fly brain, dropped on a real criminology paper, choosing its next paper from the references in front of it, stopped when it hits a paywall.
First we had to decide where the fly starts. If we picked the paper ourselves we would be measuring our own taste in papers, so we drew 100 seed papers at random and used all 100. We drew them from the Web of Science research area Criminology & Penology, taken as it comes: 51 journals, 30,654 articles.1 Under the rule in the table below, 31% of those articles are open, or 35% if you count bronze. Everything the fly knows about a paper — its title, the papers it cites, and whether it is free to read — comes from OpenAlex, a free public index of scholarly work released CC0, so the map the fly walks on is a map anyone else can download and check.
Every run then goes the same way. The fly is dropped on its seed paper, and its next move is chosen from the referenced papers. If it picks one it can read, that paper becomes where it is standing and its references become the new options. It never goes back to a paper it has already read. A run ends when nothing readable is left next to it, or at a cap of 800 papers, which exists so one lucky run cannot walk all day.
Something has to make that choice, and that is the fly. Simulating all 166,700 neurons for every step would be far too slow, so we kept the part of the brain that produces action: the 1,342 descending neurons, which are the cells that carry commands out of the brain to the body, plus two hops of their strongest upstream partners. That comes to 20,000 neurons and 1,016,133 connections, chosen out of the 151.9 million in the published file. Connections carrying fewer than five synapses are dropped, because those are the least reliable ones in the data, and each connection that stays is set to excite or to inhibit according to the neurotransmitter predicted for it; 22,764 with no usable prediction are dropped. This is a real piece of a real fly rather than the whole animal, and we say so rather than implying otherwise.
The choice itself is made by letting that circuit spike. Each neuron is a leaky integrate-and-fire unit: it adds up what arrives from the neurons feeding it, leaks a little away between arrivals, and fires when it crosses a threshold. We run 1 ms steps for 50 ms of simulated time per decision, with a membrane time constant of 20 ms. Each unread neighbouring paper is assigned one of eight bands of input neurons and stimulates it at a rate set by how often that paper has been cited. Paywalled neighbours drive a band too, so they push the circuit around without ever being a legal choice — the fly senses the wall. The descending neurons are split into eight matching readout groups, and the paper whose group fires the most wins, with ties broken by the random number generator. Gain is calibrated so the network averages 9.93 Hz, roughly a biological firing rate. Nothing learns and nothing about the fly changes between steps, and the random seed is fixed, so any run can be reproduced exactly.
OpenAlex labels each paper with how it is made free, if it is. Here is what each label means and how the fly treats it.
| OpenAlex status | Here | Why |
|---|---|---|
| diamond | corridor | fully OA journal, no author charges |
| gold | corridor | OA journal indexed in DOAJ |
| green | corridor | free copy in a repository |
| hybrid | gate | open license inside a toll journal |
| bronze | trapdoor — impassable | free today at the publisher's discretion, no open license |
| closed | wall | not free to read |
Bronze is the hard case: the publisher has put the paper up for free but granted no licence,
so it can be taken down again and cannot be reused. We treat it as impassable, which is stricter
than OpenAlex's own is_oa flag. Because that choice makes paywalls look worse, the
bronze-inclusive number is published too.
The run above is one draw from those 100 seeds.
The second video, above, shows both sides of that: on the left, spikes moving through the real wiring; on the right, the single paper the fly is standing on and every reference it senses from there, each one marked open or paywalled. Almost all of them are paywalled. That is the major finding — the fly is not clever and does not need to be, because what stops it is not the difficulty of the literature but the wall in front of it.
A different seed gives different numbers, so we ran all 100 of them. 75 of the 100 randomly drawn papers were paywalled at the seed: the fly could not take a single step, because a reader starting there has nothing to read. Of the 25 that let it start, the median run read 158 papers and bounced off about 2,000 walls — and 20 of those 25 hit our 800-paper cap, so their totals are floors, not ceilings. Across all 100 seeds it read 4,022 papers and hit 69,102 paywalls: about 17 walls for every paper it could open, roughly 5% of everything it touched. The individual numbers move from run to run. The finding does not.
Most of the fly-brain demos going around do not run a control, so there is no way to tell whether the fly's wiring is doing the work or whether anything of that size would behave the same. We ran one. The whole pipeline was rebuilt on shuffled copies of the same connectome: each neuron keeps its number of connections, but who it talks to is randomised, which destroys the fly's actual circuitry while leaving its size and shape intact. Then we measured how reliably the readout groups identify which input band was driven — if the real wiring matters, the real brain should do this better than the shuffles. One shuffle would be an anecdote, so we ran several and compared the real brain against their spread.
Do not take this too seriously as a study. It is meant to be fun, and to make a point about who can read criminology. If you want to check it, argue with it, or take it further, everything is open: the code, every trace, the corpus and this control are at github.com/crimconsortium/fruitfly. The code is released under the MIT open-source licence, and our data and figures under the open-access Creative Commons Attribution licence (CC-BY 4.0). Feedback, comments, reviews or anything else are welcome at consortium@crimrxiv.com.