Research previewMainnet not launchedView launch criteria

PROTOCOL / V0.2

From connections to proofs.

flybrain technical whitepaper · Research draft

Download Markdown ↓

01 / ABSTRACT

A connectome-driven computation experiment.

flybrain combines the open fly connectome, sparse FlyHash codes and nonce search. Miners verify the full data, build connectivity fingerprints and search data-dependent access traces for an acceptable hash.

The implementation can recompute proofs locally. This document distinguishes running mechanisms from network and proof-system research.

Open data provides a shared input. Deterministic computation provides an output that can be checked.

02 / DATA IDENTITY

Start every computation from the same data.

The input is the official MaleCNS v1.0 syn-partners table from HHMI Janelia. Each row is encoded as 40 bytes, preserving both body IDs and endpoint coordinates.

ItemImplementation
Official source file6,777,179,098 bytes
Full row count311,833,243
Canonical data12,473,329,720 bytes
Leaf size4,096 records
LicenseCC-BY 4.0

The bulk table contains many segmentation fragments. Its row and body counts are not the paper's filtered neuron and synapse counts.

Data commitment

leaf[i] = H(domain, LE64(i), canonical_chunk[i])
dataset_root = MerkleRoot(leaf[0], ..., leaf[n])

The version, canonical encoding and Merkle root identify the input. Content, rather than a download URL, determines its identity.

Locally reproduced dataset root
6841b8dd1f5892afbd9588041b656b26c18843c23f2fed9b2f042bbe9bf98c8d

03 / SPARSE REPRESENTATION

Expand, then retain the strongest connections.

Each body has 25 outgoing and 25 incoming buckets. Integer-quantized counts form a 50-dimensional vector, expanded to 2,000 outputs. Each output reads six inputs; the strongest 100 survive.

50Input dimensions2,000Sparse expansion100Winner-take-all

Output index breaks activation ties. Winners are encoded in ascending order. Integer arithmetic avoids floating-point and platform-RNG differences.

A dictionary compresses fingerprints for 88,384,522 body IDs to 2.74 GB without changing fingerprints or their artifact root.

04 / PROOF OF WORK

Every search returns to the data.

The challenge binds the template, data root, index root, epoch seed and miner identifier. Every nonce performs 256 serial accesses; each next record depends on the preceding data and hash.

  1. Derive the initial state from the challenge and nonce.
  2. Read a synapse record, its chunk hash and endpoint fingerprints.
  3. Update the state and choose the next access.
  4. Repeat 256 rounds and fold the trace into a FlyHash input.
  5. Compute the work hash and check difficulty.
challenge = H(header, dataset_root, artifact_root,
              epoch_seed, miner_key)
work_hash = H(challenge, nonce, access_trace, flyhash_code)
accept = leading_zero_bits(work_hash) ≥ difficulty_bits

Diagnostic runs support 0–20 leading-zero bits; wallet mining uses the fixed testnet difficulty. Full targets and adaptive network difficulty remain future work.

05 / VERIFICATION

From full replay to succinct proofs.

Current: full local recomputation

The service verifies the full data and builds the index. Verification replays the accesses and checks roots, challenge, trace, winners, work hash and difficulty.

Next: proofs of access

A Merkle multiproof proves membership under a root; it does not by itself prove correct construction of the full index.

Research: proofs of full computation

Chunked STARK or zkVM proofs with recursive aggregation require evaluation at the actual data scale.

Scanning and random access do not remotely prove continuous RAM residency. The implementation uses mmap, with residency managed by the operating system.

06 / SCIENTIFIC UTILITY

Utility must be measured.

FlyHash fingerprints are a candidate representation for connectivity retrieval. Scientific utility needs public queries, baselines, Recall@K, mAP, latency and construction-cost measurements.

Execution and reproducibility have been tested. Retrieval quality and the marginal value of repeated computation still need independent experiments.

07 / NETWORK DESIGN

Current scope and monetary policy.

The local coordinator now supports signed wallets, staking, attestations and a persistent test ledger. It does not provide independent P2P consensus.

Supply is capped at 311,833,243 FLY: 10% community reserve and nine mining tranches of 10% each. The first tranche targets 10,080 one-minute blocks, then each tranche doubles its block count. Rewards split 80% to miners and 20% to validators.

View roadmap ↗

08 / REFERENCES

Built on open research.

  1. HHMI Janelia · MaleCNS v1.0 data ↗
  2. Google Research · Complete male fly connectome ↗
  3. Dasgupta, Stevens & Navlakha · A neural algorithm for a fundamental computing problem, 2017 ↗
  4. Fly-LSH · MIT-licensed reference ↗