Aristotle: Long-Horizon Formal Reasoning
Nomos runs a continuous research process built around Aristotle, an autonomous agent from Harmonic designed for long-horizon formal reasoning and proof-oriented work rather than one-off answers.
Without a durable process, every promising line of reasoning has to be rediscovered from scratch, and partial progress is lost between runs. A long-horizon agent keeps what survives checking, records what fails, and returns to the same open question with more structure than it had before.
Long-horizon reasoning is an active area of published research on which progress is measured in verified steps rather than in announcements.
As autonomous systems take on longer proofs, the discipline that matters is not raw speed but the ability to hold a single objective across many cycles, revisit failed attempts, and keep every intermediate result attached to a check.
Nomos treats no mathematical claim as proven without verification. Results are separated into known, provisional, rejected and verified, and the agent’s output is published as research records rather than as conclusions.
That separation is what keeps a long-running process legible: anyone reading the public record can see which line of reasoning survived, which one failed, and what is still open.
The Agent: Aristotle
Aristotle is an autonomous research agent, built for long-horizon formal reasoning and proof-oriented work by Harmonic, the lab building mathematical superintelligence.
A long-horizon agent is judged by what it can carry forward. Every cycle either strengthens an existing line of reasoning, closes it off as rejected, or leaves it provisional until it can be independently checked or formally verified. Nothing becomes verified because it happens to look plausible.
The work runs as a loop rather than as a session. Each pass proposes a step, then subjects it to formal checking, keeps whatever survives as provisional material, and files the rest as rejected. Failed branches are not discarded quietly, because a recorded failure narrows the search space as usefully as a confirmed result does for the cycle that follows.
The engine driving those cycles is not a general-purpose chat model. It is a system built for formal reasoning and proof-oriented output, where every step is stated so that it can be machine-checked rather than accepted on the strength of fluent prose.
Because the horizon is long, the agent also has to decide what to revisit. Progress is stored so that a later cycle can reopen an earlier branch with fresh context and constraints - rather than from a blank page.
A long-horizon agent is only useful if its record is durable. Each cycle is written to the research ledger, where every entry stays marked as known, provisional, rejected or verified until something changes its status.
The current objective is deliberately concrete: reconstructing the known boundary of the Strong Goldbach Conjecture - the question of whether every even integer greater than two can be expressed as the sum of two primes.
The Objective: A Known Boundary, Rebuilt
In plain terms, the agent is not being asked to announce a proof. It is being asked to rebuild the boundary of what is already established about the conjecture, step by step, in a form that can be checked rather than simply trusted.
Why Long Horizons Matter
Short-horizon systems are good at producing an answer that reads well. Formal mathematics is unforgiving in a different way: one unjustified step invalidates everything downstream, however convincing the surrounding argument may look to a reader.
Working over long horizons changes the failure mode. The agent can spend cycles on a branch, establish that it does not close, record that outcome, and move on - so the effort is not wasted, because the ledger now holds a rejected path that no later cycle has to repeat.
Formal verification, an explicit status for every claim, and a record that survives between runs remove most of the ambiguity that unstructured research spends its time arguing about. What is left is the mathematics.
The framing is deliberately narrow. Rather than chasing an open problem end to end, the agent starts from established results, the material that has already been checked, and works outward from there.
The continuous cycle can be described in three parts:
- Test: Each candidate step is stated precisely enough to be checked, by formal verification or by an independent pass, before it is allowed to support anything built on top of it.
- Reject and preserve: Steps that fail are recorded as rejected rather than deleted, and steps that survive are preserved as provisional material that the next cycle can build on.
- Extend: Surviving lines are extended in the following cycle, so the boundary of what is known moves outward gradually instead of arriving as a single announcement.
None of this is a claim about settling the conjecture. It is a claim about method - a process that runs continuously, keeps its own record, and can be read by anyone who wants to check where the reasoning currently stands.
About Nomos & Harmonic
Nomos exists to pursue one question in public and to fund the computation that keeps the pursuit running, on the principle that truth is something continuously pursued rather than declared.
The reasoning engine is Harmonic, co-founded by Vlad Tenev and Tudor Achim, which is building mathematical superintelligence. Its Aristotle agent is designed for long-horizon formal reasoning and proof-oriented work.
Nomos publishes its research records as they are made, not after the fact. To follow the current objective, phase and ledger status, or to raise a question about the method, reach out to @usenomosmath on X.