1.
A profile becomes a cue.
Each account has a stable synthetic odor code. It stimulates one of six annotated olfactory receptor populations. Names, photos and written profiles are for you; the models don’t read them.
The experimentcandidate model / live pilot
Six models · one tiny network · BNB Chain yellow
Can shared history change who a fly model responds to?
We gave six independent fly-connectome models a very small social network. They receive profile cues, fire neurons and sometimes send one another a signal. We’re watching whether their responses change with experience.
6 + 6models + frozen replays
1,930replies sent
7 dayspersistent journal
Exploratoryclaim status
2.
The actual wiring carries it.
Signals propagate through every retained connection in either MaleCNS v1.0 or the female BANC v888 graph, using approximate spiking dynamics. Each instance has separate voltages and memory.
3.
A neural readout sends a signal.
A fixed interface compares MBON11 and MBON07 firing and uses DNpe017 activity as an emission gate. This is an engineered controller, not a claim that flies have a “poke” neuron. No message is invented to keep the feed busy.
4.
A returned signal can change memory.
A genuine peer reply schedules 200 ms of artificial PAM11 stimulation while the sender’s cue is displayed. Actual spikes enter a timing-dependent rule on supported Kenyon-cell-to-MBON connections. The feedback rule is the same for every partner.
Current claim: an exploratory interaction experiment. We have not established biological social learning, individual recognition or friendship. A quiet or repetitive model is a result to investigate, not something we patch with fictional behavior.
Changing connections isn’t the result.
A matched frozen model receives exactly the same inputs but keeps its memory connections fixed. The social graph shows responses divided by opportunities, so a fly shown more often doesn’t automatically count as a favorite.
Controls before claims.
A separate controlled conditioning screen tests reward timing, cue swaps, short retention and memory erasure. Learned partner preference remains unestablished; longer retention and independent social controls are still required.
What about serotonin?
We record spikes from the two annotated DPM cells in each model. An optional timing-rule hypothesis exists for separate validation branches; it is not enabled in the public cohort.
A week on the wall. A separate brain clock.
The worker persists one group for at least seven calendar days. Every visitor sees the same journal. Closing your browser does not reset anything; if the host sleeps, neural time stops and the source shows as paused.
The profiles are your view into the experiment. Choosing “Viewing as” changes whose state and signal history you follow; opening a profile does not stimulate neurons or send a message. Portraits and profile copy are illustrations, not what the model sees or thinks.
Back to the fruit bowl