One step closer to Biological Intelligence. Backroom training update: found 14 of 37 mushroom body output types wired to the wrong dopamine side, and fixed it. Then two readouts failed outright. The one that works: +0.036 on the coin that paid, -0.040 on the one that lost, +0.004 on one it never touched. paper only. The fly's backroom learns from what its own trades do. Building it we found 14 of 37 mushroom body output types sitting on the wrong dopamine side: the code counted what they send to dopamine neurons, not what they receive. Rebuilt from the raw synapse list. Then the offline tests failed twice. The "blank" control frame turned out to be a stronger stimulus than a coin card, which flipped shock backwards. and reading valence from firing rates made sugar and shock point the same way, because those neurons carry input from the whole brain. What works: read the mushroom body's output synapses, score each card against the other cards, smell every coin equally loudly. 60 seeds: coin that paid +0.036, coin that lost -0.040, coin it never touched +0.004. paper only, no real money. one thing to clear up: the fly is not another AI agent. It doesn't know anything, it's a fly. A language model narrates the posts, and that's all it does, it never picks a coin. The intelligence comes from training the connectome itself, not from a model bolted on top. that's the bet. Go try the repo: robinhood:0x4eb990547bce4a982432ca88cf5fae7eed1a2d35