Discovered Materials (YC P26) has developed a benchmark to evaluate AI agents' ability to discover novel, dynamically stable, and BEOL-compatible crystalline materials. The objective requires models to propose candidates meeting specific targets for thermal conductivity ($\kappa$), static dielectric constant ($\epsilon$), Young's modulus, and shear modulus, alongside a synthesis recipe that an expert would attempt. To facilitate this, models are equipped with a suite of computational tools, including web search via Exa and a Python/Bash sandbox containing pymatgen, mp_api, and ASE. Crucially, the agents utilize machine learning interatomic potentials (MLIPs), specifically the PET-MAD foundation model, to compute properties. Dynamic stability and harmonic lattice thermal conductivity are determined using Pheasy and Phonopy via compressed sensing to fit second-order force constants, while lattice thermal conductivity is estimated by solving the Boltzmann transport equation under the relaxation time approximation. Mechanical properties are derived from the compliance tensor through 12 different unit cell strain configurations. Synthesis feasibility is assessed via a rubric-based grading system—utilizing a "worst of three" GPT-5.6 Sol ensemble—which applies critical and fixable penalties to evaluate recipe quality. The benchmark allows models to iterate until they hit an error or exhaust a 100-million-token budget.