Two-minute read; links to the underlying papers, model cards and case studies inline.
We replicated the FEP+ 8-target panel against literature ground truth at 1.06 kcal/mol RMSE. Aether Bio 12B is now the default for FEP campaigns across customer deployments. SWE-bench Verified resolution rate hit 53.6%. Reproduction kits at /benchmarks.
Named aerospace prime now running structural margin signoff entirely on Aether for a transport-class airframe. Comparison memo runs four pages; redacted version on request.
Tool registration updated for the new MD-classical solver with 1.4× throughput on amber99 force field. No refusal-corpus changes this month.
How should we calibrate our refusal corpus for materials chemistry where the same precursor has benign and dual-use trajectories? Reply if your team has thought about this.
- (1) Active learning for slow ground truth — Conti et al.
- (2) Memo on long-horizon agent evaluation — Pajari.
- (3) Wind-tunnel CFD correlation under turbulent transition — internal.