Geometry quality matters: the σ-profile is a fingerprint of the molecule's actual surface, and the surface depends on where the atoms are. Reference-tier geometries are minimised until the residual gradient norm drops below 10⁻³ Hartree/Bohr — the threshold beyond which σ-profile values stop drifting at machine precision.

Tier 2 — xtb_relax=True (Industrial, planned): GFN2-xTB analytic gradient + scipy. Sub-second per-drug. For high-throughput screening where DFT geometry would be overkill.

Convergence

Each .mfsig.json records the optimised geometry, the converged gradient norm, the number of optimisation steps, the methodology used, and the SHA-256 of the canonical state. Re-runs on the same hardware reproduce bit-identical output.

When it helps

Geometry quality matters most for flexible scaffolds where the σ-profile depends on torsions and dihedrals. For rigid aromatics it matters less, but the Platinum tier delivers Platinum-grade geometry on every drug — no per-molecule decisions for the customer.