The treasury is the mechanism, not the model
There is a version of this project that would have been much easier to explain. Publish weights, publish the data manifest, ask for donations, and let the thing run until the donations stop. Every part of that is already common practice, and every part of it has the same ending.
The codex names the problem in the manifesto rather than in a chapter, which is unusual for a document this dry: a model without continuous capital is a candle without oxygen. The real problem of AI is not intelligence. It is endurance.
What actually funds a forward pass
Validators are paid from the treasury, per attested job. The treasury is refilled by the swap fees earned by a liquidity position that the treasury itself owns and holds. That is the whole loop, and it has four stations: the token is traded, the protocol-owned position accrues fees, those fees land in the treasury, the treasury pays the validators who answered queries.
Two properties fall out of that shape, and both matter more than they sound.
Nothing is minted into the cycle. Supply is fixed at 1,000,000,000 at genesis and the mint authority is revoked. Validator rewards are therefore bounded by revenue, not by a schedule: the protocol cannot issue an obligation it cannot meet, because it has no instrument with which to issue one. This is design principle P8, stated as a constraint rather than an intention.
Nothing leaks out of it. The token contract carries no transfer tax. Third-party holders are not charged for moving the asset; only the protocol's own position participates in the revenue cycle. A transfer tax would have been the easier mechanism and it would have made every holder a payer. The token chapter chose the harder one.
Why this is the part with no precedent
Open weights are not novel. Distributed inference is not novel. What is novel, and therefore the part most likely to be wrong, is funding continuous operation out of the asset's own trading activity rather than out of a balance sheet someone had to raise.
The codex builds the floor into the mechanism. A reserve floor, expressed as a minimum treasury balance in months of trailing compute spend, blocks disbursement below it, and a timelock sits in front of every disbursement above it. The treasury cannot be drained faster than the people relying on it can see it happening.
That is the trade. A network funded by a boardroom fails when the boardroom loses interest. A network funded by its own liquidity fails when nobody wants the asset. We think the second failure is the one worth having, because it is the one the people using the network can actually see coming.