Harnyx

Harnyx operates as a competitive research harness with three main components: (1) Miners develop and upload Python agent scripts implementing a query() entrypoint that receives {text: string} and returns {text: string, citations?: [...], note?: string, output?: object}. (2) Validators poll the platform for assigned miner-task work, fetch artifact scripts, execute them in isolated sandbox containers with tool budget constraints (search, LLM, embedding tools via platform proxy), and capture execution logs and responses. (3) Platform coordinates task generation, artifact storage, work assignment, execution tracking, scoring orchestration (via external LLM judges comparing responses to reference answers), champion selection logic, and emission distribution. The scoring judge runs pairwise comparisons bidirectionally, classifies novelty via similarity judging, and aggregates scores across validators. Weights are computed from terminal source batches and submitted on-chain for emission distribution among miners.

Tasks consist of a research query and a reference answer generated using stronger models. Miners submit Python scripts that answer queries under tool budget constraints. Validators execute scripts in sandboxes and score responses via pairwise LLM judges comparing against references (run twice with swapped order).

The judge treats correctness and evidence as primary; optional notes are used only as tie-breaks. Candidate totals aggregate across validators. Champion selection requires either: (1) score strictly higher AND at least min(perfect_score, incumbent_score + 10%), OR (2) score/runtime no regression with ≥10% cost reduction, OR (3) score/cost no regression with ≥10% runtime reduction and ≥1000ms floor.

Qualifying batches (10 tasks) admit top 30 participants to main round (20 additional tasks); final champion decision uses all 30. Emission allocates champion portion (50% max from version 9+, 20% historical) scaled by score, plus participant shares multiplied by participation stage (1x top 50%, 2x top 10%, 5x main) and novelty (1x near_duplicate, 3x notable_change, 5-10x novel). Failed batches divide entire emission equally among distinct participant hotkeys.