How Rat Processing Unit works

7,126 NFTs on Robinhood Chain. Each one is a single neuron recorded from a real rat’s brain. Holders fire their neurons into one shared, simulated rat cortex, and every result is written on-chain so anyone can re-run it and check.

The neurons

The recordings come from a published dataset: 128-channel silicon probes placed in the neocortex of 20 rats under ketamine/xylazine anesthesia. The researchers sorted the signals into 7,126 single units, which are individual neurons as best as spike sorting can tell. That count is the supply.

Each token carries its neuron’s real measurements:

• Spike shape: the neuron’s average spike, 81 samples at 20 kHz (about 4 ms). Stored fully on-chain after the reveal.
• Firing record: every spike it fired during its recording session, at the original 1/20,000 s precision. Hosted off-chain and locked by fingerprint.
• Traits: firing rate, burst index, spike width (narrow or broad), amplitude, cortical area, estimated layer, and the rat it came from.

Rarity comes from these measurements, shown as percentiles across all 7,126 neurons. Nothing is generated or invented. Two caveats: the layer is the researchers’ estimate from the probe position, and “narrow” vs “broad” spike width is our labelling (under or over 0.45 ms).

Traits and rarity

Every trait is derived from the neuron’s real measurements by fixed rules, so anyone can recompute them from the published data. Nothing is random or hand-picked.

Rarity score adds up how surprising each of its traits is (−log₂ of how common it is), plus a bonus for titles. Neurons are then ranked from #1 (rarest) to #7,126; every rank is unique. Tiers follow the rank: Legendary (top 1%), Epic (top 5%), Rare (top 20%), Uncommon (top 50%), Common.

The fair mint

1. Before anyone mints, the list of all 7,126 neurons is shuffled and its fingerprint (provenanceHash) is set on the contract. It locks at the first mint and can never change.
2. Everyone mints blind. Nobody knows which neuron a token will get, including us.
3. When the mint ends, the contract picks a future block. Its hash sets a random startingOffset. Token N gets the neuron at position (N + startingOffset) mod 7126 of the shuffled list.
4. Anyone can finalize the reveal, including from the mint panel on the site, so nobody can stall it or pick a better outcome.
5. After the reveal we publish the shuffled list. Anyone can check it matches the fingerprint that was locked before the mint.

Mint fees go to the team’s wallet. The team can also mint before the reveal; those tokens are assigned by the same random reveal as everyone else’s.

The shared cortex

The cortex is the Blue Brain Project’s model of rat somatosensory cortex: 5,924 neurons across six layers and 568,717 synapses, built by EPFL from real measurements (at 7% of real neuron density). We run a simplified version in your browser:

• leaky integrate-and-fire neurons, with connection strengths and delays taken from the model’s synapses;
• a simple short-term depression rule, so activity stays bounded;
• integer arithmetic only, so every computer gets exactly the same result.

When you fire, each of your neurons drives 24 model cells in its estimated layer. The busiest stretch of its real recording is replayed, spike for spike, as input. This mapping is our design choice, not biology. The site draws the model’s response layer by layer.

Holder features

• Name your neuron. Owners can give their neuron a name of 1–24 letters, digits, spaces, dashes, apostrophes or dots. Names are unique (ignoring letter case), stored on-chain, and stay with the neuron when it is sold; the new owner can rename it. The project can remove an abusive name but can never set one.
• Early access to new brains. When a new cortex version is added, holders use it first on the site for a week. Everyone else keeps firing the previous version until then. Every recorded fire notes which brain version it used, so old results always verify against the brain they ran on.

Verify it yourself

Every recorded fire stores two fingerprints: an input digest (which brain, which neurons, how long) and a result digest (every spike the model produced). The brain files are locked by their own fingerprint on the cortex contract.

In the browser

Under “Recent fires on-chain”, press Verify. Your browser downloads the brain, checks its fingerprint, re-runs the fire, and compares both digests with the on-chain record.

From the command line

# 1. get the public files (brain + neurons) from the project's data host
# 2. with Node.js 20+ in that folder:
node fire.mjs <neuron> [<neuron> …] <milliseconds>
# prints inputDigest and rasterDigest: compare with getFire(id) on the cortex contract

Neuron numbers for tokens come from neuronOf(tokenId) on the NFT contract plus the published shuffled list. The brain fingerprint rule is sha256(sha256(cells.bin) ‖ sha256(conn.bin) ‖ sha256(params.json)).

What is on-chain, and what isn’t

Contracts

Three contracts on Robinhood Chain. Addresses appear in the site footer at launch.

• RPUNeurons: the NFTs, the fair mint and the delayed reveal. Uses Arbitrum’s ArbSys for block numbers and hashes, as Robinhood Chain is built on Arbitrum.
• RPUCortex: brain versions (append-only; a fix adds a new version and old records keep theirs) and fire records.
• RPUWaveforms: every spike shape, stored fully on-chain, writable only after the reveal, then frozen.
• RPUNames: names for neurons, set by their owners and stored on-chain.

Settings such as the mint price stay adjustable. Metadata, the data root and the spike-shape store each have a one-way freeze. The contracts are covered by 42 automated tests, including fuzz tests and invariant tests run across over a hundred thousand random calls.

Data and credits

• Recordings: Horváth C, Tóth LF, Ulbert I, Fiáth R (2021). Dataset of cortical activity recorded with high spatial resolution from anesthetized rats. Scientific Data 8:180. CC-BY.
• Cortex model: Blue Brain Project, EPFL. SSCx toy circuit with 6k neurons in SONATA format, doi:10.5281/zenodo.12202781. CC-BY 4.0.
• Brain anatomy: Waxholm Space atlas of the Sprague Dawley rat brain v4 (Papp et al. 2014, NeuroImage 97; Kleven et al. 2023). CC BY 4.0. Surface meshes from MeshView, Neural Systems Laboratory, University of Oslo.

Rat Processing Unit is not affiliated with or endorsed by these researchers, EPFL, the University of Oslo, Robinhood or Amazon.

FAQ