Robinhood Chain · mainnet 4663
An AI with a smart contract for a brain.
Wired compiles neural networks into Robinhood Chain contracts. Every inference runs onchain, and every neuron can be checked.
- Block
- —
- Model
- 0x0fdd…f6d1
- Gas
- —
- Cost if sent
- —
- Local replay
- —
Live on mainnet · registry 0x320d…7ca7 · weights 0x059a…54ae
How it runs
No server. No oracle. The contract does the math.
Compile.
Train in PyTorch or scikit-learn. Wired turns the weights into int8 and checks that the integer model answers like the float one.
- Shape
- 64 → 48 → 32 → 10
- Weights
- 5,018
- Float accuracy
- 95.6%
- int8 accuracy
- 95.6%
- Same answer
- 100%
- Test digits
- 450
Store.
The weights become blobs: contracts with no logic, only bytes. A model is the hash of its blobs, so it can never change under you.
Byte 0 is STOP, so the blob can never run. The rest is read with EXTCODECOPY.
Run.
One 256-bit word holds 32 weights. Split into lanes, a single multiplication moves 8 neurons forward.
Numbers
A million parameters fit in one transaction.
| Model | Parameters | Gas per inference | Cost | Onchain output |
|---|---|---|---|---|
64→64→10 minimal game bot | 4,810 | 156,946 | 0.000003 ETH | identical to reference ✓ |
128→128→64→8 arena bot brain | 25,288 | 416,663 | 0.000008 ETH | identical to reference ✓ |
784→128→64→10 handwritten digits, 28×28 | 109,386 | 1,203,694 | 0.000024 ETH | identical to reference ✓ |
256→512→256→16 ~0.26 M parameters | 267,024 | 2,438,291 | 0.000049 ETH | identical to reference ✓ |
512→512→512→512→16 ~0.8 M parameters | 796,176 | 6,130,692 | 0.000123 ETH | identical to reference ✓ |
1024→768→512→256→16 ~1.3 M parameters | 1,316,368 | 10,040,727 | 0.000201 ETH | identical to reference ✓ |
Measured with eth_call on Robinhood Chain mainnet, November 20, 2026, gas price 0.020 gwei. Random int8 weights; each output compared byte for byte with the reference engine.
Build
Any contract can ask a model.
A model on Wired is a view call away. Gate a mint on a drawing, let a game bot decide its next move, score a trade before it settles. The answer is computed in the same transaction, by the same chain.
- infer(id, input)logits, free as a view call
- classify(id, input)label and logits
- decide(id, input)same, written to the logs
- register(blobs, name, uri)content-addressed model id
interface IWired {
function classify(bytes32 model, bytes calldata input)
external view returns (uint256 label, int256[] memory logits);
}
contract SevenLock {
IWired constant WIRED = IWired(0x320D078b873912cf8dDe00f750fd16240C457ca7);
bytes32 constant DIGITS = 0x0fddc7aa3584b716fb13b93fd278b55f0a93205d1a745ba0bfbc195c64c3f6d1;
// opens only for someone who can draw a 7
function unlock(bytes calldata drawing) external view returns (bool) {
(uint256 digit, ) = WIRED.classify(DIGITS, drawing);
return digit == 7;
}
}Runtime, registry and arena are live on Robinhood Chain: registry, runtime, arena. From any language, use the HTTP API.