HKUST Paper Shows Reversing JSON Field Order Drops Agent Success From 85% to 19.2%

A new HKUST paper argues trained agent policies secretly encode their training harness's action scores, and small presentation changes can wreck performance. Reversing MiniWoB++ JSON field order leaves every optimal decision unchanged yet drops task success from 85.0% to 19.2%; on Terminal-Bench Lite, frozen Qwen3.5-9B ranges 47.9%-64.1% across five harnesses. The authors propose a C(I)-M(θ,I) ceiling-minus-mismatch decomposition to pick harnesses jointly with training.