Industrial cohorts demand consistent per-drug wall time. The reference 100-drug Platinum cohort delivered in 38.6 minutes; the 389-drug DrugCentral cohort runs in 2.5 hours — same audit trail per drug, no quality variance.

MINAO init guess saves ~5 SCF cycles vs the legacy 'atom' SAD default. Coarse-to-fine two-pass SCF saves 20-40 % on drug-sized molecules. Schwarz screening at 1e-9 (was implicit 1e-12) throws away more near-zero integrals. preset='fast_gga' bundles all of it with r2SCAN (no exact exchange).

Per-knob tuning

omp_threads cap, init_guess (minao default), grid_level + grid_level_coarse, schwarz_tol, level_shift, diis_space, auxbasis. All captured in the .mfsig audit trail for reproducibility.

Cohort throughput

Cap each worker to 8 OMP threads and run n_cores / 8 workers in parallel via multiprocessing / Ray. 20× wall reduction at cohort scale, zero algorithmic change.