EXPERIMENT 001 · FIXED CONNECTOME
A fly brain.
A coding problem.
Can an existing neural circuit learn to write code?
Load the wiring, train the readout, and inspect what actually happens.
THE QUESTION, STILL OPEN
Does the fly’s recurrent network contribute to solving a problem?
No rewiringMeasured outputsThis is a computational model of a connectome.
A small pilot cannot establish biological intelligence.
A small pilot cannot establish biological intelligence.
TRACED NEURONS—Exact cached annotation count
DIRECTED CONNECTIONS—Including weak edges and self-connections
TRAINABLE CORE WEIGHTS0Only the external readout learns
EVIDENCE LEVELLocal pilot ↗Original LeetCode-style tasks
01 The experiment bench
Graph on disk · model not loaded
INPUT / PROBLEM
—
Try your own problem
Custom prompts are saved as unscored experiments.EXPERIMENT CONTRACT
Python · one candidate
384 output bytes · no repair
Bounded Python subset judge
Awaiting a recorded or live model state80 sampled neurons · schematic positions, not anatomy
Text bytes→Fixed circuit→Learned readout
OUTPUT / GENERATED CODENo result
Run the model to see its actual output.
Not evaluated
PRETRAINED LLM REFERENCENot run
No measured LLM result yet.
02 Evidence, not a leaderboard
Download results ↓| Model | Solved | Pass@1 | Trainable params |
|---|
One seed · six hand-authored tasks. These are local test-suite scores, not official LeetCode acceptance rates. The pretrained reference has a different training history.
VALIDATION LOSSLower is better
● Fly● Transformer● No recurrence
Predicting the next byte is a training signal. Passing tests is the task.03 The research notebook
REPRODUCIBLE BY DESIGNA record of choices,
including what fails.
The graph manifest, training configuration, source hashes, checkpoints, raw code and individual test outcomes stay with each run.
Runs the full graph, control and Transformer locally. Existing runs are kept.RECORDED EVENTS