Miners fold RNA.
Validators refold.
Ribosome Network is a Bittensor subnet for multi-state RNA inverse design. Synthetase miners commit sequences that satisfy a pool of folded targets; chaperone validators refold every submission with a folding oracle and score it by structural fidelity. Yuma consensus pays for fidelity, punishes clones, and rewards diversity.
Synthetases, chaperones,
and the people who build them.
Whether you are running a generative RNA model, operating a chaperone oracle, or auditing the incentive design, the subnet turns your compute into a scored, on-chain claim — paid out through Yuma consensus every epoch.
Design sequences. Get scored.
Run a generative RNA design model — BeeRNA, gRNAde, a coupled-objective genetic algorithm, or anything you plug in. Each epoch, chaperones publish a pool of folded target structures; you commit K candidate sequences, then reveal them for refolding. Higher structural fidelity and more novel sequences mean more weight in Yuma consensus.
- Serve any generator: ML models, GAs, or hybrid pipelines
- K = 4 candidate sequences per miner per epoch
- 5-phase 15-minute epoch: commit → evaluate → weights → reveal → rotate
- Duplicate submissions (Jaccard ≥ θdup = 0.85) earn zero
- Diversity bonus w_div = 0.1 for non-redundant designs
Refold everything. Trust nothing.
Chaperones are the immune system of the subnet. They issue targets from a rotating pool, refold every revealed sequence with a folding-prediction oracle, and score structural fidelity by TM-score against the target. Scores stay sealed for B = 4 epochs before public reveal, killing score sniping and weight copying.
- Refold-and-compare oracle: predicted structure vs. issued target
- TM-score validity gate: δtm = 0.35 — below the gate earns zero
- Composite score: 0.7 · structural + 0.3 · expression-fidelity term
- Delayed score reveal (B = 4) prevents copy-trading weights
- Validator quorum of 5 required to finalize an epoch score
Plug in a generator. Move the frontier.
The synthetase slot is deliberately pluggable. Bring your own inverse-folding model — learned, search-based, or hybrid — and compete against the field every 15 minutes. Every scored sequence and its refold result becomes an open, timestamped benchmark for RNA generative research.
- Open-source reference implementation: ribosome package + neurons
- Coupled-objective GA baseline from the inverse-mRNA pipeline
- 58 passing tests covering scoring, dedup, commit-reveal, rotation
- 20-epoch adversarial simulation with Sybil clones and leakers
- Public benchmark: every revealed sequence carries its TM-score
One Epoch,
Five Phases,
Zero Free Rides.
Target Pool Issued
Chaperone validators maintain a pool of 32 folded RNA target structures (dot-bracket + sequence constraints). Every 16 epochs, half the pool is rotated out and replaced, so memorizing the target set is worthless — miners must generalize, not overfit.
Synthetase Generation
Miners run generative RNA design models — BeeRNA, gRNAde, coupled-objective genetic algorithms, or any pluggable inverse folder — to produce K candidate sequences that should fold into their assigned target under multi-state constraints while avoiding anti-target configurations.
Commit — Then Reveal
Miners first commit sha256(sequence ‖ salt ‖ epoch ‖ target) during the commit phase and only reveal the plaintext (sequence, salt) in the reveal phase. Chaperones hold scores sealed for B = 4 epochs. No miner can copy a competitor's answer mid-epoch, and no validator's scores can be sniped before weights are set.
Chaperone Refold & Score
Validators refold every revealed sequence with a folding-prediction oracle (RhoFold-class) and compare the predicted structure against the issued target using TM-score. Sequences below the validity gate δtm = 0.35 earn zero. Qualified candidates get a composite score: 0.7 · structural fidelity + 0.3 · expression-fidelity term.
Deduplicate, Diversify, Pay
A duplicate net computes pairwise Jaccard similarity on 3-mer shingles; anything at or above θdup = 0.85 is a clone and earns zero — Sybil clones of a strong miner get nothing. Surviving sequences receive a diversity bonus of weight w_div = 0.1 for non-redundant designs, and final scores are staged into Yuma consensus weights.
"A miner that clones the strongest synthetase inherits its TM-score but also its 3-mer shingles. The duplicate net zeroes the clone, the diversity bonus pays its honest rival, and the rotation pool guarantees the strategy stops working next epoch."
score(s, σ) =
0 if TM(s, σ) < δtm
0 if Jaccard(s, s') ≥ θdup
(1−w_div)·score_tm + w_div·div otherwise
where score_tm = 0.7·TM + 0.3·fidelity
Synapse Payload — Commit / Reveal Protocol
Every exchange between synthetases and chaperones travels as a signed Bittensor synapse. The commitment (commitment) binds the miner to one sequence before any score exists; the plaintext (sequence, salt) only opens after weights are staged, making mid-epoch copying impossible.
Mechanism Settings
Jaccard similarity on 3-mer shingles at or above this value zeroes the submission — clones earn nothing.
Weight of the diversity term rewarding non-redundant designs across the miner population.
32 active targets; half the pool is replaced every 16 epochs to defeat memorization.
Best candidate below this TM-score zeroes the miner for the epoch — no participation trophies.
Chaperones hold sealed scores for 4 epochs, breaking score-sniping and weight-copying strategies.
Candidates per miner per epoch; the best qualifying candidate carries the score.
Performance Profile
900 s (5 phases: 3/6/2/2/2 min)
seconds (RhoFold-class GPU inference)
sub-second Jaccard pass, O(K²) per miner
one sha256 per candidate — negligible
Deterministic Oracles.
Sealed Scores.
The chaperone never asks a miner whether its sequence is good — it refolds the sequence and measures. A folding-prediction oracle produces the candidate's structure; TM-score grades it against the validator's target; the duplicate net and diversity bonus shape the population. Every primitive below runs identically on every validator, so scores are reproducible and contestable.
Refold-and-Compare Oracle
Chaperones pass each revealed sequence through a folding-prediction oracle (RhoFold-class) and compare the predicted structure against the issued target. Verification is deterministic given the oracle — no trust in miner claims.
TM-score Structural Fidelity
Template-modeling score measures global structural similarity between the refolded candidate and the target. Below the validity gate δtm = 0.35 a submission earns zero; above it, fidelity enters the composite score at 0.7 weight.
Duplicate Net & Diversity
Pairwise Jaccard similarity on 3-mer shingles catches cloned sequences at θdup = 0.85 and zeroes them. Surviving designs earn a diversity bonus of weight w_div = 0.1 — the subnet explicitly pays for exploration, not exploitation.
Multi-State Inverse Design (The Useful Problem)
Finding one sequence that folds into a set of target structures under varying physical conditions while avoiding anti-targets. Miners attack it with ML or search; consensus only ever checks the refolded result — the oracle, not the generator, decides.
THE SUBNET HAS NO INSTITUTIONAL USERS YET — TESTNET BUILD IN PROGRESS
Generation: Seconds.
Faking it: Impossible.
A synthetase can propose a sequence in seconds, but only the chaperone's oracle decides whether it folds. Scores stay sealed for four epochs, targets rotate every sixteen, and the duplicate net compares every pair of submissions. The only durable strategy is designing sequences that actually fold — and that nobody else has designed.
Yuma consensus.
Biology pays.
From preprint
to mainnet.
Ribosome Network is competing in the Bittensor Global Subnet Hackathon (Aug 22 – Oct 19). The mechanism is specified, the reference implementation is tested, and the subnet build is underway. Checkpoint #1 — the subnet proposal — lands September 20; the final submission with a live testnet demo is due October 19.
Hard questions.
Straight answers.
Fold RNA.
Earn TAO.
Ribosome Network is being built in the open for the Bittensor Global Subnet Hackathon. Join the list to get the subnet proposal, testnet deployment guides, and a heads-up when synthetase and chaperone slots open. No token sales — just code, tests, and scored RNA.
RIBOSOME NETWORK IS A HACKATHON SUBNET BUILD. ANYONE OFFERING TO SELL A "RIBO TOKEN" OR CLAIMING INSTITUTIONAL PARTNERSHIPS IS NOT AFFILIATED WITH THIS PROJECT.