Progress you can
inspect for yourself.
A useful discovery comes with evidence. Inspect submissions, reproduce settlement calculations, and check scored checkpoints.
Every settlement is recomputable from the published record.
The validator publishes one JSON record per round: revealed sources, commit hashes, the baseline it scored against, every held-out loss, the checkpoints with their SHA-256 and the weights it set. The verifier runs the validator's own code on those files. A mismatch means the validator did not publish what it computed.
Commit and reveal
Each source hashes to its commitment. Each scored commit has a public preimage. Task and data pins are stable from round to round.
Clean scoring
Quartile checkpoints at 25, 50, 75 and 100 percent carry a held-out loss and a hash. Download the blob and recompute the loss on your own GPU.
Bar, events, weights
The adaptive bar, the frontier outcome and the weights are replayed from the prior rounds at the round's recorded timestamp and compared to what was published.
# replay every settlement from the round files python -m sn125.verify --rounds-dir rounds/ # fetch the checkpoints and check their hashes, then recompute the held-out loss python -m sn125.verify --fetch-base YOUR_ARTIFACT_BASE_URL --rescore --device cuda
What cannot be recomputed by a third party is the training run itself: production uses non-deterministic kernels for throughput. The checkpoint trajectory, the clean-scored loss, the hardware provenance recorded per score and the validator's signature on every record are the audit artifact; the source is public and anyone with a B200 can run it again.
Credit follows the contribution.
Published frontier optimizers are attributed to the submitting hotkey. A signed authorship claim lets contributors connect the work to a public identity, or remain pseudonymous.