AsymmetryZero: A Framework for Operationalizing Human Expert Preferences as Semantic Evals

AsymmetryZero (arXiv:2605.04083) introduces a structured framework converting subjective human expert preferences into explicit semantic evaluation contracts, defining grading criteria, judgment logic, and aggregation rules for RL post-training pipelines. Contracts execute via Inspect for model-only evals and Harbor Framework for agentic settings, producing comparable scores and shared audit artifacts. Teams building reward signals for domain-specific tasks, where exact-match or open-ended preference judgments fail, should prioritize this immediately.