EXPERIMENT 001 / A NEW BODY

One brain.The first learned jump.

A connectome-derived circuit controls an unfamiliar body. Observe signals becoming actions.

Live spike simulation · In your browser1,024 -cell subcircuit, not a complete fly brain.
Only the action decoder learns; neural connection weights are frozen.

BODY 01 / Jumper

READY · Awaiting experiment

NO INPUTGAME 0.00s / BRAIN 0ms

Reading connectivity data…Manual: ← → / Space to jump

Progress0.0 %

Body horizontal speed0 px/s

Body stateGrounded

Total reward0.00

Training & frozen evaluation

TEACH → TRAIN → TEST

An explicit teacher demonstrates right and jump. Train the spike-to-key mapping, then remove the teacher for evaluation.

Untrained · Initial decoder00 / 12

Teacher action cross-entropy—

Decoder weight change L20.000

Neural weight change0 · Frozen

Condition	Wins / Trials	Mean progress
Before training	Not evaluated	—
After training	Not evaluated	—
Sensory input disconnected	Not evaluated	—
Random keys	Not evaluated	—

Fixed evaluation seeds 101–105, ; these are short-level tests only. The decoder also reads stimulated cells, so failure without input does not prove internal connections are essential.

Take the experiment with you

SNAPSHOT / EVIDENCE

Save the neural state, decoder, body and evaluation record. Import at the same paused instant to continue or branch.

↑ Import & validate

No record hash yet

Robinhood Chain · Experiment commitments

Register the protocol and completed evaluation hashes. Requires a wallet and test ETH. The chain stores hashes; neurons run in the browser.

No transaction submitted. Testnet 46630.

Who chooses the next body?

Holders may eventually choose experiments and compute budgets. No token, treasury, governance contracts or trading fees are live.

Methods, full-network evidence & sources

Loading the recorded local full-network experiment…

Browser pipeline: four public game-state features → artificial stimulation of 21 sugar sensory cells → 1,024 -cell LIF subcircuit → 16 spike readouts → four-action decoder → game physics. This is not biological vision; the controller receives no full map.

Training: a rule teacher labels the same local features. Twelve levels, class-balanced softmax cross-entropy SGD. Only 68 external decoder weights change.
Reward: new maximum progress / 10 − per step 0.01, goal +100, fall −20. ; imitation training uses no reward gradient. Reward is an evaluation metric only.
Browser integration: 1 ms Euler steps, 24 ms of neural time per 1/30 s of game time; the original 1.8 ms ms delay is rounded to 2 ms.
Selection: start from 21 source sugarR cells and select two hops by outgoing contact count up to 1,024 cells, retaining the induced graph. Layout is not anatomy; removed connections are absent.
Full-network records are independent 200 ms CPU smoke tests, not a browser livestream or statistical replication. The undefined reset assignment w=0 was removed; other default dynamics were retained.

Connectivity & cell IDs · Full-network spike record · Original model · MIT License · Anatomy credits