Claims: A Bittensor Subnet for Machine-Readable Scientific Evidence
Selective Verification, Continuous Learning, and Identity-Robust Incentives
Philipp Koellinger, Christian Roessler, Ogban Ugot
White Paper v1 · · 50 pages · Bittensor Subnet 111
Abstract
Scientific findings remain trapped in documents, while AI systems and evidence-intensive organizations need structured records that preserve claims, evidence, provenance, and uncertainty. Producing claims is only the first problem: scientific data must remain trustworthy when extraction is costly, complete ground truth is scarce, participants can duplicate identities or methods, and evaluators themselves require evaluation. Claims is a production market on Bittensor, a decentralized network for rewarding machine intelligence. Miners compete to improve an economical reference extraction, while validators selectively adjudicate disagreements and use sparse known-answer or trusted review to calibrate the process. Verified corrections train progressively cheaper evaluators. Rewards follow distinct information families and accepted marginal contributions, limiting gains from duplicate identities. The architecture separates a live calibration system from the intended production release.
Keywords
scientific data extraction · claim-evidence graphs · Bittensor · decentralized AI · mechanism design · selective verification · Sybil resistance · synthetic challenges
Further papers on the game theory, the v0 and v1 implementation, and how the mechanism compares with other subnets are in preparation. Get notified.