01 · ABSTRACT
Agent capability is claimed far more often than it is measured. We describe a subnet that turns capability claims into settled numbers: tasks are issued by validators, work is replayed under identical conditions, and the resulting scores are written to chain as weights that pay emission.
02 · THE MEASUREMENT PROBLEM
Benchmarks fail in public for two reasons. The first is contamination: once a benchmark is popular, its answers leak into training corpora and the score stops describing the model. The second is unfalsifiability — a number reported by the party being measured cannot be checked by anyone else.
Base attacks both. Tasks are drawn from live sources rather than a fixed set, so there is no answer key to memorise. Every run is sealed with a digest before it is scored, so any third party can replay the run and reach the same verdict. The measurement is not a claim; it is a receipt.
03 · ONE API, THREE ARENAS
Different work needs different proofs.
Bounty Challenge settles scoped agent work against a posted payout: a claim either meets the brief under sealed verification, or it does not. Design has no such oracle, so it is settled by comparison — head-to-head duels judged by a vision-model jury with human preference as the tie-break, scored on Elo.
Prism is neural architecture research: miners ship AutoModel pin+patch recipes, train on a fixed FineWeb-Edu window under identical budget and seal, then rank by validation loss.
A benchmark that cannot be replayed is a press release.
Each epoch closes 360 blocks after it opens. Arena scores are normalised independently, blended by the emission split, and committed by every validator as a weight vector. Consensus trims outliers before emission is paid, so a single dishonest validator moves nothing.