Research

We let the chain mark the homework.

Most model training asks another model to score the answer. We run the transaction on a private copy of Solana and look at what it actually did. A plan either moved the money it claimed to move, or it did not.

01 Corpus

Transactions, paired with what they did.

The training set is not documentation and not forum posts. It is landed transactions matched to the state they produced, and failed ones matched to the error that explains why.

Mean coverage across every epoch

37.9%

1,064 epochs185 complete86 missing

Slots0 to 459,653,000StatusReplaying backwards from the cutoff

One cell per epoch, 1,064 of them, from genesis to slot 459,653,000. Replay runs backwards from the cutoff, so the recent chain is whole and the deep history is not. Red cells are ranges no archive node we have access to can serve. The dataset card publishes this same table, gaps included.

02 Anneal

Trained against a copy of the chain.

We fork a validator at a known slot, let the model act, and replay what happened. The reward is the difference in on-chain state, so there is no judge model to fool and no style points for a confident answer.

03 Soliton Bench

Seven tests, graded by replay.

Nothing here is multiple choice. A task passes when the transaction lands and the result matches what was asked for. One suite scores a model for refusing, because confidently signing the wrong thing is the expensive failure.

Tasks

3,890

Across seven suites.

Suites

Writing, reading, diagnosing, routing, budgeting, forensics, refusing.

What is not done.

This is the list we would want to read about someone else. It is maintained as things change, and it is linked from every number on this site that is a target rather than a result.

S1, S1 Pro and S1 Verify have no published eval run.: Every latency and accuracy figure on this site for those three is a target we are holding ourselves to, not a measurement. The model cards will say so until a run replaces them.
Corpus is indexed to slot 459,653,000 and stops there.: Replay of the full ledger history before epoch 1064 is in progress. Coverage per epoch is published in the dataset card, including the gaps.
Rotor has not shipped.: Half of Alpenglow is still ahead of the network. Our latency budget assumes today's propagation, and it will need redrawing when Rotor lands.
The fork grader is not adversarial yet.: It replays at a pinned slot with historical order flow. It does not yet model a searcher reacting to our agent. Until it does, route scores read optimistic.
Random-reward controls are not finished.: Published work in 2026 showed spurious rewards lifting some base models as much as real ones. We will not publish an RLVR gain without the control beside it.

Rotor has not shipped.

Alpenglow landed in two halves and only the first one is live. Our whole latency argument assumes how the network moves data today, and it will need redrawing when the second half arrives.

Checked 29 October 2026.1

Rotor status and the Alpenglow rollout: orbitflare.com, read 29 October 2026. Network behaviour changes, and this page is only as current as that date.
On spurious rewards: published work in 2026 showed random reward signals lifting some base models nearly as much as real ones, which is why no gain is published here without its control. Spurious Rewards, arXiv.