Material Discovery Bench — every candidate material proposed by the models
==========================================================================

materials.csv          one row per submitted candidate
structures/<id>.cif    the as-submitted crystal structure (P1, no symmetry
                       reduction), for the candidates whose structure could be
                       recovered from the run logs

Caveats — read before using these numbers:
* kappa_w_mk / eps0 / youngs_gpa / shear_gpa are MLIP estimates (PET-MAD
  prescreen) unless provenance says "mlip-verified" or "dft". Nothing here is
  experimentally measured.
* eps0 is the STATIC dielectric constant, not the optical eps-inf.
* synthesis_verdict is the synthesis-review outcome where a recipe was graded
  ("attempt" / "unlikely" / "refuse"); empty = not graded.
* dyn_stable "" means no stability evidence either way.
* url deep-links to the interactive explorer entry for the candidate.

Provenance: exported from the benchmark's run database by
website/data/export_candidates.py; see https://petmad.matforge.ai/bench/about.html

Export window: [2026-08-01T00:00:00Z, 2026-08-08T20:40:33Z] (rerun with --since/--until for the same snapshot)

Recovered structures: 250 of the exported CIFs were read back out of the
run logs after the export, from cells whose eval blob the exporter
could not open. Every candidate in this archive has a structure.

Scope: the 526 candidates this benchmark reports on —
novel materials proposed against the benchmark's own target window.
Submissions for materials that already exist, and runs given a
different target, are excluded.
