A Claude-backed hydrologic-modelling agent calibrates the dCREST model on the training gauges of a split and predicts daily discharge in held-out, anonymised basins it never saw. Google's Harness² refines the agent's own harness (prompts, skills, scripts) between attempts. The baseline is dCREST's own differentiable optimizer: a parameter-learning network trained by gradient descent through the model on the same training gauges. Everything is scored on the 1989–1999 benchmark test period with USGS observations neither side sees.
Loading…
Calibration framing: one climate region held out
More optimization rounds: k=10 sequential with Claude Sonnet 5.5
Which role matters? Opus vs Sonnet as Harness² editor, 3 runs per harness
Gemini 4 Argon as the agent (Antigravity CLI), Opus 5.5 as editor, 3 runs per harness
Differentiable baseline across splits
Median test NSE on the held-out basins of every split. Random folds hold out basins interspersed with the training gauges; leave-region-out splits hold out a whole NOAA climate region. The reference column transfers the released per-basin calibrations by attribute similarity; the agent only ran on the split marked with ●.
Transfer-only pilot (agent given the released donor calibrations)
The first experiment: the agent received the released per-basin parameters of the donors and only had to transfer them to 30 basins of a random fold.
Median test NSE by method
Dot = median across basins; bar = 25th–75th percentile. Left: all 665 basins, 5-fold spatial cross-validation. Right: the 30-target pilot task the agent runs on.
Distribution of test NSE
Cumulative share of basins below each NSE (pilot targets; values clipped at −1). Curves further right are better.
Leaderboard (pilot task)
Median test NSE / KGE over the pilot targets. The agent must beat the best classical regionalization (attribute similarity) to count as an improvement; the gauged upper bound uses each target's own calibration and is not reachable without data.
Classical baselines by climate region (all 665 basins)
Median test NSE per NOAA climate region, 5-fold spatial cross-validation.
Agent rollouts
Every executor rollout of the Harness² run: the base harness, the probe candidates, and the composed harness selected without seeing any score.