A single episode on the Classical Mechanics environment. A Gemini agent is given three bodies in a bounded 2D box and a fixed measurement budget. Each observation costs budget depending on how many bodies it tracks and at what fidelity. The agent must iteratively decide what to measure and when, use a Python sandbox to test hypotheses about the force law, then derive the equations of motion and predict future positions — all before the budget runs out. The true law here is altered gravity (force ∝ 1/r), so memorized inverse-square intuition does not apply.
The agent spent its budget mostly on high-fidelity full-system snapshots, leaving little room for the long-horizon measurements needed to pin down the force law.