Minos — subnet 107

Minos turns a genomics benchmark into a repeated contest, and its published scoring code is unusually specific about exactly how a narrow win becomes an outsized reward.

What it does

Minos runs a genomic variant-calling benchmark. Roughly every seventy-two minutes the platform behind it generates a fresh challenge genome containing hidden synthetic mutations, injected at the level of individual sequencing reads by a tool called HelixForge. [source] A miner never uploads a result — it downloads that challenge, runs one of three established variant-calling tools (GATK, DeepVariant, or bcftools) with its own chosen settings, and submits only the settings used, called a hyperparameter config. [source]

A validator then re-executes that same config itself against the same challenge genome, and scores the result against the true, known mutations using hap.py, an industry-standard genomics comparison tool. [source] Nothing an operator reports is taken on trust — its own re-run of that same submission is what actually gets scored. [source]

How the incentive mechanism works

The repository lays out where each piece of this lives: neurons/miner.py and neurons/validator.py are the two chain-facing loops, utils/scoring.py runs hap.py inside Docker and computes the score, and utils/weight_tracking.py turns a round's scores into the weights actually submitted on chain. [source]

The score itself, called the Advanced Score, blends four components on a 0–100 scale: the core F1 match rate is worth 60% of it, completeness 15%, the false-positive rate 15%, and a quality measure the remaining 10%. [source] [source] That score is divided by 100 to become a single round score between zero and one, and only the current round's score counts — nothing from earlier rounds carries over into it. [source]

Turning scores into pay is deliberately winner-heavy: the single best-scoring eligible operator in a round takes roughly 90% of that round's weight, the next nineteen split the remaining tenth on a sliding scale that favours whoever is closest to first, and nothing currently is withheld through burning. [source] Eligibility itself requires having scored in at least five of the last twenty rounds, including the current one — fall short and pay stops until that bar is cleared again. [source]

What miners are actually optimising

The stated goal is accurate variant calling. What the reward structure actually pays for is narrower: being the single best config on that specific round's synthetic genome, because the gap between first place and everyone else is worth roughly nine times as much as the entire rest of the field combined. [source] [source]

Unverified With only three underlying tools available and a config limited to their published parameters, a miner's real edge over the field may come down to how thoroughly it has searched that tool's parameter space against this platform's own mutation-generation habits, rather than to any deeper advance in variant calling. The repository's own documentation gestures at this directly — describing today's contest as hyperparameter search, with custom algorithms saved for "once the hyperparameter space has been saturated" — but this dossier cannot confirm how much headroom is actually left in that search space today, or how close to saturated it already is. [source]

Why it earns

Under this network's own design, a contest's share of a day's pay against every other contest runs on one formula and nothing else: share_i = p_i × (1 − b_i) / Σ_j p_j × (1 − b_j), where p is that contest's own moving-average price and b is a penalty for pay an owner withheld to its own keys. [source] [source] A well-specified, checkable scoring formula like Minos' own can only move that number by first moving price — by making enough people want to hold the token — so a documented mechanism and a documented price are never fully separable from measurements like this one.

Mechanism evidence

What this dossier can confirm is that the mechanism is unusually specific and checkable: a published four-component formula, an industry-standard comparison tool doing the actual grading, and a judge that re-runs an operator's own submitted settings rather than trusting a report. [source] [source] That specificity is evidence the mechanism is legible, not evidence that legibility is what is driving the pay described below — the two remain separate questions this snapshot cannot join.

Concentration evidence

Minos carries 2695 @ 8871732 holders and a comparatively deep root proportion of 35.1% @ 8871732 — a large share of the judging power behind it traces back to money staked generally into the root network, rather than staked specifically into Minos' own token. [source] A contest whose judging leans this heavily on generalist backers is being watched substantially by people who did not necessarily pick this specific genomics benchmark on its own merits.

Price evidence

Minos' token rose 7.47053686181305 @ 8871732 percent over the trailing week and 35.71071348147201 @ 8871732 percent over the trailing month, against a price of 0.064803558 @ 8871732 TAO and 3260654638738 @ 8871732 rao traded over the preceding day. [source] Because the cross-contest formula above answers only to price, a rising price over exactly the stretch this snapshot captures is, by construction, a real part of why Minos' own emission share sits where it does — not merely a coincidence alongside its mechanism.

Failure modes and attack surface

Paying roughly 90% of a round to a single winner is a design that rewards a hair's-width margin as if it were a decisive one, which raises the value of squeezing out the very last bit of a known tool's parameter space over genuinely improving variant calling more broadly. [source] The same structure means an operator who occasionally lands first can carry a whole stretch of otherwise-average rounds, so a small number of very sharp submissions can outweigh a much larger number of solid but unremarkable ones.

The round's ground truth — which mutations are real — is generated and served entirely by Minos' own platform, not by anything a validator can independently check against a second source. [source] Re-running a submitted config keeps individual operators honest against that truth, but nothing pinned here lets this dossier check the truth itself, or the mutation-generation process that produces it, for bias or predictability round over round.

What to copy, what to avoid

Worth copying: a validator that re-executes an operator's own declared method against fresh, unseen data rather than grading a self-reported result — it closes off an entire category of dishonest reporting for free. [source]

Worth avoiding: concentrating close to the entire round's reward on a single winner. It is a clean, legible rule, but it also makes the difference between a genuinely better method and a slightly luckier hyperparameter search look identical from the outside, and it puts a large share of a contest's total pay at the mercy of whichever single submission happens to land first each round. [source]