Cards arrive one at a time, and for each one the question is: is this the same odor as the card just before it? The three flies' answers are the real answers the step 10 model gave. The real brain is a model built from one male fruit fly's wiring diagram (the MaleCNS connectome); the shuffled brain keeps every neuron and every connection count but rewires who connects to whom; memory, no brain sends the olfactory receptor neurons' signals straight into the memory unit, bypassing the brain. Switch on Play along to compete with the flies.
The benchmark. 1,000 items that show the flies 33 odors they never smelled in training, through olfactory receptor neurons never used in training (test B of step 10 in the repository). Each fly's "Benchmark" score is over all 1,000 items; "This deck" is over the 160 items in the deck being played.
How to read it. The real-brain fly tells same from different even when a new odor arrives on neurons it never used. The shuffled-brain fly and the memory-without-a-brain fly are at chance even on the cards they trained on — they memorised instead of learning — so this game could not decide whether the wiring is what makes the difference (VOID). Comparing the two cards' brain responses by cosine similarity alone, with no memory, the real brain separates same from different odors at AUC 0.947 and the shuffled brain at 0.514; that was explored after the results, so it is a hypothesis for a next experiment, not evidence.
What is real and what is staged. Every comparison is one real benchmark item: the model received the previous card's odor on one set of receptor neurons and, after one blank frame, this card's odor on neurons unused in training, and answered (mean probability over three initialisation seeds, split at 0.5). The continuous stream is stitched together from items whose second odor matches the next item's first odor, so the model drew fresh neurons and started a fresh memory for every comparison. The shuffled-brain fly is the first shuffle by number (seed 1000), not picked by score. The bars on each card are the responses of 23 receptors (up excitation, down inhibition); a fly's input spreads that pattern over randomly chosen neurons, so it is not identical to the bars. The flies, the table and the cards are illustration.
Sources and licenses. Receptor responses: Hallem & Carlson (2006) Cell 125:143, via DoOR.data v2.0.1 (Münch & Galizia 2016, Sci Rep 6:21841), CC BY-SA 4.0; the response patterns on these cards are shared under the same license. Connectome: MaleCNS v1.0 (Berg et al. 2026, Cell), CC-BY 4.0. Code and rules: step 10 of the malecns-operator repository.