DGrid AI's latest research introduces a Proof of Quality (PoQ) framework to address a core flaw in decentralized AI scoring, specifically the reliance on having a correct answer to compare against for node reward distribution. The paper, the fourth in DGrid’s ongoing research series, proposes using small evaluator models to score each output’s quality without needing ground truth data. DGrid built this framework with a cost-aware version that bakes latency into payout math, an adversarial-robustness layer, and a granular breakdown of “quality.” The research notes that an off-the-shelf NLI cross-encoder returned a Pearson correlation of −0.363 when used to rate answer quality without a reference answer, meaning it was more likely to favor poor responses. Instead of adapting existing models, the researchers trained specialized AI judges to score output quality, improving decentralized AI reward systems at scale.