Would you merge code reviewed by a fruit fly?

The nose

Each receptor above 25% drives its glomerulus at probability × 150 Hz. Identifiers and operators from the changed lines are hashed onto the remaining 23 glomeruli, with the background of keywords and punctuation adapted out. The odor selects the 5% of Kenyon cells with the most weighted input from those glomeruli’s projection neurons in the real wiring; everything downstream is simulated. Jev is never asked for a verdict and never learns what the fly did.

The measured output

What is real and what is ours

The point cloud is measured neuron-body positions. Bright points are a sample of neurons that fired in the displayed 25 ms window. The six larger ringed cells are the readout neurons; the small rings are MBONs (red avoidance, green approach) and dopamine neurons (pink). Camera movement is not activity.

Lines are sampled real connections drawn straight between cell bodies, not axon shapes. The incoming signal rail is a diagram. The simulation uses all 176,422 neurons and 6,287,749 retained connections; the picture is downsampled. Each trial starts from rest. Given the same odor, seed, and memory, a trial is bit-identical (Rerun the exact saved inputs checks this); the only thing carried between trials is the memory file.

Ours: the thirty questions, which glomerulus each drives, the token hashing, the 5% Kenyon-cell selection standing in for APL inhibition, the ×5 on KC → MBON synapses that lets the odor code drive the MBONs, the form and rate of the plasticity rule (a synapse from a recently active Kenyon cell weakens while the dopamine neurons wired onto its MBON fire), the drive that tops up the dopamine neurons because this model’s taste pathways into them are incomplete, the 0.45 and 0.68 lines, the 8 Hz reach line, and the words “changes requested” and “no objection”. Calibrated by presenting odors directly, never on the examples. The fly’s: which Kenyon cells an odor reaches, which dopamine neurons reach which MBON and so which synapses a bitter or a sweet taste can change, which MBONs count as approach or avoidance (the compartments PPL1 innervates against the ones PAM innervates), what its MBONs do afterwards, and whether its proboscis motor neurons fire. The score is not a probability of a bug, and a connectome that learns is not a connectome that understands.

Data: male CNS v1.0, by HHMI Janelia, Cambridge / MRC LMB, and Google Research, CC BY 4.0. Connections below five synapses and autapses were removed by the upstream derivative.

The verification

Run an independent investigation to generate a test.

A changed result only counts as a regression when the original satisfies the written contract. No counterexample in the executed cases is not proof of correctness. Pasted changes have no contract, so they get a vote and no test.

The memory

Learning happens inside the run. A lesson is a taste delivered while the odor is on: bitter on the proboscis after a reproduced regression, sugar after a designated behavior-preserving control the sweep could not fault. The fly’s own dopamine neurons fire (PPL1 for bitter, PAM for sugar), and every 5 ms a synapse from a recently active Kenyon cell is weakened by however much dopamine the connectome delivers to its MBON. Which synapses can change is the wiring’s decision; the amount is our rate. “No regression found” on anything else stays unknown. The vote is behavior first: if the proboscis motor neurons fire to the smell alone, the fly reached for it. Otherwise it is the balance of avoidance and approach MBONs, and a naive fly sits in the middle and abstains.