Hyvmind: A Research to Earn dApp for Tokenising Annotation
Anshuman SinghWhile the superiority of language models over traditional search-engines has become clear in recent years, their ‘generativity’ remains a serious concern for policymakers. On the input side, it is generally acknowledged that most large models are trained on unethically sourced data. And on the output side, their tendency to hallucinate and misinform makes them unfit for domain-specific work. This paper takes the view that explainability and transparency cannot be achieved simply by putting ‘a human in the loop’. Taking legal work as a concrete site, it proposes Hyvmind – an architecture that puts ‘humans in the centre’ by recording and rewarding semantic labour through tokenised annotations. Its novelty lies in conceptualising legal research as a set of four
interconnected functions (source, watch, frame and curate) around a common data-object (source-text). By storing and rewarding annotative-work through a distributed ledger system with nested states, it creates a secure, ethical and organic pathway for generating high-quality datasets for the next generation of domain-specific language models.