Fees Become Thought: Funding Continuous Inference
A share of Nomos token tax revenue is reserved for computation. That reserve funds the model inference behind the research process, which is what allows the work to run continuously - testing, rejecting, preserving and extending - rather than in isolated, one-off bursts of activity.
Research meant to run for a long time carries a mundane requirement underneath it: someone has to pay for the compute. Reasoning over long horizons is not a single query, and its cost does not arrive once. It recurs with every cycle, which is why funding is treated as part of the design rather than as an afterthought, and why Nomos reserves a share of protocol tax revenue for exactly that purpose.
The loop is deliberately short. Token activity contributes to the reserve, the reserve pays for inference, inference produces research records, and those records are published as they are made. Nothing in that sequence waits on a funding round or a grant cycle arriving on time.
The distinction matters more than it first appears. A process funded in discrete grants runs until the money for that run is gone, then stops, and whatever context it had built up is lost with it. A process funded by a standing reserve does not have to stop, which is what makes a long horizon possible at all.
What the spending buys is not a product. It is inference - the model computation that lets the agent propose a step, check it, discard what fails and carry the rest into the next cycle, repeatedly, without a person deciding when the next run may begin.
Fees become thought: activity in the token funds the reserve, and the reserve funds the reasoning that produces the record.
Where the Money Goes
The chain is short enough to state in full. It begins with the protocol tax, the input: a share of token activity is directed into the project's compute reserve instead of being spent elsewhere.
The second link is the reserve itself, the research budget. Funds held there are allocated to sustained machine reasoning and to nothing else, which is the entire point of keeping the reserve separate.
The third link is the process: Harmonic, co-founded by Vlad Tenev and Tudor Achim, whose Aristotle agent is built for long-horizon formal reasoning and proof-oriented work.
The fourth is the output - the research ledger, where known, provisional, rejected and verified results are kept distinct from one another. In outline:
- Input and reserve: token activity contributes to the compute reserve, and the reserve is allocated to sustained machine reasoning rather than to general operating costs.
- Process: the reserve pays for the inference that keeps the agent testing, rejecting, preserving and extending its research over time instead of in single runs.
- Output: the record that spending produces is public, and known, provisional, rejected and verified entries remain distinct within it rather than being blended into a summary.
Why Continuity Changes the Work
The case for funding compute this way is methodological rather than financial. The framing the project uses for it is short - a research process that has to stop is a research process that forgets - and the reserve exists to prevent that:
A process that pauses between funded runs loses the state it had built: the branches it had ruled out, the partial results it was carrying, the reasons a particular direction looked promising in the first place. Continuity is not a luxury here. It is the thing that lets a long-horizon agent treat its own earlier work as material rather than as history.
That is the practical reason a share of tax revenue is reserved rather than left to be allocated later, when the case for continuing to spend would have to be made again.
The reserve's purpose is narrow by design: it pays for inference. It is not a treasury for general spending, and it is not held back for a launch that would make the research look more finished than it is.
What the Reserve Buys
What continuous funding buys is a horizon rather than a headline.
Because the compute is paid for ahead of any particular result, the agent is free to spend cycles on branches that turn out to close nothing. Those cycles are not waste under this model: they produce rejected entries, and a rejected entry narrows the search for every cycle that follows it.
That trade is only available to a process which does not have to justify each run before it is allowed to begin.
The reserve does not make a result more likely to be correct. It makes it possible for the work to continue long enough for correctness to be established one check at a time.
Recorded in the research ledger, alongside the objective and phase it belongs to.
No mathematical claim is treated as proven without verification, and no amount of compute spent on it changes that.
Spending, in other words, is not evidence. The reserve determines how long the process can run and how many cycles it can afford; it says nothing at all about whether a particular entry in the record deserves to be moved out of provisional.
The record keeps those two things apart, which is why activity counts - research cycles completed, tokens processed - are reported next to the ledger rather than in place of it.
Nomos & the Compute Reserve
Nomos exists to pursue one question in public: whether every even integer greater than two can be expressed as the sum of two primes. The current objective is narrower than the conjecture itself - reconstructing its known boundary - and the reserve is what keeps that objective funded from one cycle to the next.
The economic loop supporting it has four stages, each of which does exactly one thing:
- Input - protocol tax
- Reserve - research budget
- Process - Harmonic
- Output - research ledger
Harmonic, co-founded by Vlad Tenev and Tudor Achim, is building mathematical superintelligence, and its Aristotle agent is designed for long-horizon formal reasoning and proof-oriented work. The reserve pays for the inference that agent consumes, and the ledger records what that inference produced.
The ledger is being initialized and no verified result has been recorded yet, which is what the record should say at this stage rather than something more encouraging. To follow the objective, the phase, and the entries as they are written, reach out to @usenomosmath on X.