Surrogate science is the methodology of developing, assessing and validating proxies for high stakes decision making, where the proxy is a substitute for gold standard measurements of success that are too expensive, time-consuming, long-horizon or unavailable to obtain. In biomedical research, therapeutic evaluation tools like biomarkers, preclinical models, new alternative methodologies (NAMs), digital medical twins, intermediate clinical endpoints are all putative surrogates. In businesses, KPIs for eventual business success are a surrogacy problem. In economics, short-term indices for long-horizon outcomes are a surrogacy problem. In AI evaluations, LLM graders and LLM judges as a proxy for human evaluations are a surrogacy problem. In AI safety, model organisms are a surrogate for evaluating efficacy of AI safety interventions, evaluating the faithfulness of chain-of-thought as a proxy for internal computations is a surrogacy problem.

The goal of the surrogate science project is to

• Disseminate the science of developing good surrogates and evaluating surrogacy for high stakes decisions, since surrogacy is a species of causation not correlation.
• Demonstrate that many high stakes decision making problems can benefit from explicitly acknowledging surrogacy issues
• Learn from the frontiers of biostatistics, econometrics and causal machine learning that have been at the forefront of solving surrogacy problems.
• Champion investment in surrogate science R&D across all fields of science, engineering and policy