Episode 3 of 10 · 15–16 Sep 2026 · 4 min read

The fly that turned only left

We spent two experiments hunting a broken neuron in the map. The broken part was ours.

By Denis

Simulation onlyNo hardware in this episode.

A grid of small pictures: the scene from a camera as the fly walks past three coloured pillars, and the same scene as the fly's many-lens eyes see it, blurry and pixelated — Simulation · what the fly seesWalking past three pillars: the plain camera view, and the same moments through his 721-lens eyes.

Stripes

Flies steer by what they see. Show a fly stripes sliding past to one side, and it turns to follow them. It’s one of the oldest experiments in fly science, and a good test of whether our brain works the way a real one does.

Real · video from YouTubeA real fruit fly in a virtual-reality arena, walking on a floating ball and steering by what it sees. YouTube loads only when you press play. Video: “Neural Compass”, SciTech Daily

So our fly got eyes, built like a real fly’s: 721 tiny lenses on each side. Then we showed him moving stripes.

He turned left.

Stripes the other way: left again. Whatever the stripes did, he turned left.

The suspect

The obvious suspect was the map itself. Real scientific maps have gaps, and one steering cell, called DNb06, looked lopsided: 7,030 connections coming in on one side of the brain and only 749 on the other, with a note in the data that part of the cell was missing.

A damaged cell on one side would explain a fly that only turns one way. It was a good story. Two experiments went after it.

The first switched the suspect off. That removed about 40% of the one-sided signal: real, but not the whole answer.

The second tried to “repair” the cell, and we tested the repair against a sham: a fake repair that shouldn’t change anything. The repair didn’t beat the sham. So we couldn’t say it worked, however much we wanted to.

The twist

Then we followed the signal step by step, from the eyes to the legs. The main turning cell, DNa02, had been answering to both directions the whole time. Left stripes, left signal. Right stripes, right signal.

The brain was fine. The bug was in our own hand-written formula, the one that turns brain signals into leg movements. It had been reading the brain lopsidedly.

Once we read DNa02 directly, he turned both ways: 125 degrees a second one way, 134 the other.

He also got better at his actual job: he found and ate the sugar 85.7% of the time, up from 63.2%.

Reading the right neuron, he turns both ways, and eats more

Turning speed, degrees a second

Before, any stripesleft

After, one way125 L

After, other way134 R

Found and ate the sugar

Before63.2%

After85.7%

Chart · simulated runsBefore the fix he turned left at 109 to 153 degrees a second, whichever way the stripes moved.

What we kept

Two habits came out of this week, and they run the lab now:

Write the pass-or-fail rule before you look. We borrowed this from medical trials, where it’s called pre-registration. Before a trial starts, the team writes down what it will measure and what counts as success, so a disappointing result can’t be quietly re-scored afterwards.
We do the same for every test. The plan and its rule are saved and fingerprinted before the run, so nobody, human or AI, can move the goalposts after seeing the result.
Always run a sham. It’s our version of a placebo: a fake treatment next to the real one, the only way to tell a real effect from wishful thinking.

When the result says the famous map has a flaw, check your own work first.