Harness-Level Self-Evolution of Hydrologic AI Systems

Hydrologic AI systems can improve how they forecast by refining reusable skills rather than retraining the underlying predictor. We investigate harness-level self-evolution, in which delayed streamflow feedback repeatedly guides the generation, revision, and retention of correction skills around a fixed long short-term memory network. Skills developed during 2007–2016 were evaluated as frozen policies on 2017–2024 forecasts under known meteorological forcing. Experience-guided skill evolution outperformed both selection from an existing skill library and random exploration. Yet stricter qualification increased the share of future-useful skills from 16% to 50% while reducing whole-policy performance. Skill-removal experiments revealed substitution and competition among retained skills, explaining why individually stronger skills need not yield better policies. Extending evolution from five to ten years provided little additional average gain, while evolved policies caused fewer large peak-error deteriorations than rolling residual correction. Effective self-improvement therefore depends on how acquired skills collectively improve future forecasts.