The mouse said no

I have spent a lot of time telling people that a wiring diagram of the brain is the way to understand it. On October 16, 2026 we tested that claim on the one piece of mammal brain where it can be tested, and the answer was no. Not yet, and not with a map alone.

Here is the test, the result, and why I think it is good news.

The one cubic millimeter

The MICrONS project mapped one cubic millimeter of mouse visual cortex: about 200,000 cells, every synapse between them, finished in 2025 by a $100 million program. For 12,894 of those neurons we also know what they did in the living animal, because the mouse watched movies and gratings under a microscope before the tissue was cut. For each one we know which orientation of a bar made it fire.

So the question is sharp. Given a neuron's complete wiring, every input and every output, can you predict which orientation it prefers?

How we tested it

We froze the design before the first run. 3,000 cells, chosen by a fixed rule across layers and areas, 1,309 of them orientation-selective. Six ways of predicting each cell's preferred orientation, scored by the median error in degrees on the selective cells. Here's how it works:

1. A base rate. Guess the typical orientation for that cell's layer and area.
2. A location rule. Average the orientations of the 20 nearest cells whose tuning we know.
3. The published wiring rule. Average the orientations of the cells that synapse onto it, weighted by synapse count. This is the like-to-like rule from the MICrONS papers.
4. A fitted rule that combines location and wiring, with weights fitted on a separate 1,000 cells. This is the best mechanical rule on the whole map.
5. An agent with the whole map. An AI agent with tools to query the cell's inputs, outputs, neighbors and their tuning.
6. An agent with the wiring only. Same agent, no positions, no neighbors.

Plus a control: a model told only the cell's layer, area and class, with no map, to check whether anything was being recalled from memory.

The pass condition for the primary test was the agent beating the fitted rule by 3 degrees with an interval excluding zero. 9,000 agent runs, all on subscription accounts, $0 paid.

What happened

The agent with the whole map scored 25.8 degrees. The fitted rule scored 25.3. The difference is +0.5 degrees with an interval of -1.6 to +3.0. Fail.

Look closer and the table says something stronger. The published wiring rule, 32.9, is at the base rate, 32.8. Adding wiring to location moved the fitted rule by half a degree. The agent with the whole map landed exactly on the location rule. The agent restricted to the wiring did beat the mechanical wiring rule, by 3.5 degrees, because it weighted strong inputs and discounted inhibitory ones. It still finished 4 degrees behind knowing where the cell sits.

We saw the same thing in a first pass over all 12,894 cells before the agents ran. The published rule, with at least 10 identified inputs, predicted orientation at 34.2 degrees of error. The 20 nearest cells did it at 29.3. Replacing a cell's real inputs with random cells at the same distance did about as well as the real inputs.

The memory control came back clean. The names-only model sat at 34.8 degrees, above the base rate. Nothing about these neurons was being recalled. Whatever the agent did, it did from the map.

Failed as expected, and informative

I want to be careful about what this does and does not say.

It does not say that wiring does not matter. Of course it does. It says that in a map with one snapshot of activity and no way to run an experiment, a cortical neuron's function is read from where it sits, not from the wiring you can read off the map. Everything that saw positions landed at the location rule, agent or script alike. Everything that saw only wiring landed between the base rate and the location rule.

That said, there are caveats. 95 percent of a cell's input synapses in this volume come from axon fragments with no identity, and most identified axons are not proofread. We checked the 705 cells with at least 5 inputs from proofread axons and got the same answer, so the result does not rest on segmentation quality alone. But nobody has tested a complete, proofread input map, because none exists. We tested one function, orientation. We used one agent, one prompt, one model.

We had 6,796 more eligible cells ready to run. We did not run them. The interval on 1,309 cells is already tight enough that a 3-degree advantage would have shown up, and spending another 11 hours of compute to tighten a null was not worth it.

Why this is good news

Think of a city map. It shows every road and every building. It does not tell you what happens inside the buildings at 3 p.m. on a Tuesday. For that you need to watch, or better, to walk in and ask.

A wiring diagram is the city map. On the fly, where a simulation exists that lets you drive and silence any neuron, our agents found real mechanisms and corrected the simulation's mistakes. On the mouse, there is no simulation. There is a map and one afternoon's worth of activity. The agent cannot ask the neuron what it does, so it guesses from the neighborhood, and a 20-line rule does the same.

The gap between a connectome and an understanding is the model you build on it. That gap is now measured rather than argued, and it is why we are starting with the fly, where the model exists, and why building the model is the work.

If you have a view on how to turn a static cortical map into something that runs, I want to hear it!

The full write-up, with every interval, the pre-registration and the per-cell outputs, is at braininsights.ai.

Dan