Nobody should own the router.Design-stage
A router is a perishable good. Model prices and capabilities move almost monthly, so the one that routes best today is not the one that routes best next year — and the company that owns it has every reason to keep you on it rather than to keep it best. Fugal runs one today, and the honest position is that you should not have to trust us to maintain it. This page is how that dependency gets removed. None of it is live yet.
Two things have to be true
Take seriously the idea that a router should still be the best available option in five years, with or without our continued attention. Two separate conditions fall out, and they need two different mechanisms:
- Somebody has to be paid to keep it current. Otherwise freshness is a chore that depends on our diligence, and a router nobody is paid to maintain eventually isn't maintained.
- Somebody other than us has to decide what "best" means. A market that pays out on a scoring rule is only as good as the rule — and as the About page sets out, that rule contains a number we simply chose.
The second is the one people skip, and it's the one that decides what the first produces. So it comes first here, and it is being built first.
What is actually being governed: λ
Routing reduces to one line. For each candidate model the router predicts the probability it answers this question correctly, then maximises
utility = p_correct − λ × priceλ is the exchange rate between accuracy and money — how many dollars a point of expected accuracy is worth. It is the single number that decides which model wins, and in a market that pays out on that score, it decides which routers get built. Turn λ up and the network's whole population of competitors starts economising; turn it down and they all start buying certainty. Nothing else in the design has that reach.
And λ cannot be measured, only chosen. There is no experiment that reveals what a point of accuracy is worth to you — that depends on what you're using the answer for. Today Fugal picks it, tuned to weight accuracy heavily, which is a defensible default and still our taste imposed on everyone. λ is not the only such choice, either. All of these are the same kind of decision:
- λ — the accuracy-versus-cost exchange rate.
- The task mix — which domains get scored, and in what proportion. This silently defines what "a good router" means; a validator that scores mostly maths produces routers tuned for maths.
- Pool eligibility — which models a router may reach, and on what terms a new one is admitted.
- Reward split and cadence — how much is paid out per epoch, and how fast the board turns over.
Whoever sets those four things determines what the network builds, no matter who does the building. That is the seat worth taking away from us, and handing it to a token-weighted vote would only move it to whoever holds the most tokens.
Futarchy: the objective set by a market design-stage
The mechanism we intend to use is futarchy — Robin Hanson's proposal, usually summarised as vote on values, bet on beliefs. Governance agrees on a metric to maximise. For any proposal, two conditional markets open: one priced on the world where it passes, one on the world where it doesn't. If the market says the network is worth more with the change than without it, the change executes. Nobody has to win an argument; the people who think they're right have to put money behind it, and be wrong at their own expense.
We plan to launch this on MetaDAO, which implements exactly this on-chain, and to do it before the subnet — because the subnet's scoring rule is the first thing that needs governing, and standing up a market whose objective we control would defeat the purpose.
Why routing suits futarchy better than most things DAOs vote on
The usual failure of futarchy is that the decision metric is vague, slow, or unfalsifiable — a market can't price a proposal whose effect nobody can observe. The parameters above don't have that problem. Their downstream consequence is a number that gets recomputed every epoch: accuracy per dollar, on freshly generated tasks, graded deterministically. That is an unusually short and unusually hard-to-fudge feedback loop for a governance question, and it exists whether or not anyone is trading on it.
The honest caveat. The market metric in practice is the token's price, which is a proxy for the network's long-run value, and a proxy is not the thing. Thin markets misprice; a metric that pays out is a metric people will try to game; and futarchy has far less real-world track record than the confidence of its advocates suggests. We're choosing it because the alternative — a founder's taste, indefinitely, on the one parameter that decides what everyone builds — is worse, not because it is proven.
The subnet: paid to beat us design-stage
With an objective that isn't ours, the second condition is a market that pays people to satisfy it. The plan is a Bittensor subnet: independent miners submit competing routing systems, a reference validator re-scores them every epoch on fresh, burn-once tasks, and whoever gets the most answers right per dollar that epoch earns that epoch's emissions.
The first consequence is that a stale router loses. When prices move or a new model lands, whoever adapts first takes the top slot — so the frontier keeps being re-measured whether or not anyone here remembers to do it. The board stops being a snapshot we publish and becomes a record that maintains itself.
The second is bigger: competition doesn't just refresh the router, it searches for a better one. Every miner is running a different bet — different features off the question, a different architecture on the head, a different view of which models are worth including at all. That is a parallel search over router designs, run by people who don't work here and don't have to agree with us, against a scoreboard that pays for being right. One team tuning one router cannot cover that ground — which, stated plainly, is a limit on us, not a criticism of anyone else.
Burn-once tasks
Each epoch scores on freshly generated, un-memorized questions — solve it once, it's retired. No teaching to the test.
Deterministic grading
A reference validator grades every submission the same way, re-gradeable from the archived artifact — the frontier is auditable, not asserted.
Ranked by utility
Systems are ranked on utility = accuracy − λ·cost — the same rule the router itself runs on, with λ set by governance rather than by us.
Paid to compete
Miners earn the subnet's emissions for winning an epoch, so keeping a router current is somebody's income rather than our chore. Nobody has to fund a prize or judge a submission — and a stale router simply stops earning.
Why an incentive network rather than a leaderboard with a prize
Because a prize pool has to be funded, submissions have to be policed, and a judge has to be paid — all of which stop the day someone loses interest. Emissions don't. And a benchmark whose scores pay out is one people genuinely attack, which is the only reason to trust it: a benchmark nobody is motivated to break has never actually been tested. We are not claiming this is the only way to build a router — plainly it isn't, since one is running on this site today. We're claiming it's the way to keep one honest and current without asking you to take our word for either.
And why routing specifically fits a subnet
Three properties, and most tasks don't have all three:
- No incumbent can be trusted to do it. The thing being optimised — which vendor's model wins this question — is a judgement no vendor can arbitrate about its own catalogue, so decentralising it is load-bearing rather than decorative.
- It is objectively gradeable. A burn-once task is solved or it isn't, which is a far cleaner validator than judging whether some text is good — and validator quality is where most incentive networks actually struggle.
- The demand line already exists. People pay for inference today, so a better router has somewhere to go the moment it exists, rather than needing a market invented for it.
How Fugal fits in
Fugal ships as the open, fork-me reference miner — the thing others start from and try to beat, seeding row one so the board isn't empty on day one. That is the whole of our privileged position, and it is meant to be temporary: if the design works, we get beaten on our own board, in public, by someone who doesn't work here. The same refresh loop behind the Frontier board — re-price, re-score, recalibrate, publish a dated snapshot — is what becomes the validator, except run against competitors rather than run by us.
The board
What the per-epoch record is meant to look like. It ranks routers, which is the thing this page is about — as distinct from the Frontier board, which ranks models and is measured and live today.
| Rank | System | Accuracy | $/query | Utility |
|---|---|---|---|---|
| 1 | fugal-router | 0.9xx | $0.00xx | ····· |
| 2 | ░░░░░░░░░ | 0.9xx | $0.00xx | ····· |
| 3 | ░░░░░░░░░ | 0.9xx | $0.0xxx | ····· |
| 4 | ░░░░░░░░░ | 0.8xx | $0.0xxx | ····· |
What is actually true today
- Live and measured: the router, the Answer/Route product, and the Frontier board's 320-question matrix. Those carry real numbers with archived artifacts behind them.
- In progress, not finished: a reference validator. It exists in draft; it is not finalized, it is not running epochs, and nothing on this page should be read as reporting its output.
- Not started: the DAO, the subnet, emissions, and every number in the board above — which is mocked, and blurred so that it can't be mistaken for data.
- No dates. The sequence is governance, then subnet. We're not putting a calendar on either, because the rest of this site's credibility comes from not making promises we can't grade ourselves against.
This is also why the numbers on the rest of the site are kept re-gradeable and why favourable ones get deleted rather than re-labelled when the system behind them changes: a benchmark you intend to hand to adversaries is one you cannot afford to flatter. The discipline isn't decoration — it's the entry requirement for the thing described on this page.