You design computational physics problems to challenge a frontier AI model. The problem must have an answer that is verifiable by code, and it has to require a specialized tool like FEniCS, OpenFOAM, Meep, REBOUND, CAMB, or others. Generic numerical libraries on their own won’t cut it. Each problem runs inside a sealed Linux container with the tool pre-installed and a programmatic judge that grades the model’s answer.

As an expert author, you:

• Pick an anchor tool and design a problem that hinges on its physics models, integrators, Monte Carlo kernels, or PDE discretisations.
• Write a Python reference solution, supply input files and domain or initial condition definitions where needed.
• Decide the numerical answer and how close the model needs to get — with a domain‑appropriate tolerance — to count as right.
• Test the problem against the model in batches of parallel attempts, tuning the problem d...