Notes from the lab.

No results yet — this is the evaluation methodology we're building toward: composition-disjoint splits across twelve property classes, so a model is graded on genuinely new chemistry, not lookalikes of its training set.

No timings to report yet — this is the architecture we're designing so a candidate screen returns while you're still typing, not after a coffee break.

The design space and the constraints that actually bind — and, separately, what we're building to search it. No Accretion screening results yet; the company is pre-proof-of-concept.

A field roundup, not a benchmark post: what DeepMind, Microsoft Research, and a new wave of AI-native startups have actually shipped, and what it means for anyone entering the space now.